Slide1: BUDT 725: Models and Applications in Operations Research by Bruce L. Golden R.H. Smith School of Business Volume 1- Customer Relationship Management Through Data Mining


Customer Relationship ManagementThrough Data Mining: Customer Relationship Management Through Data Mining Introduction to Customer Relationship Management (CRM) Introduction to Data Mining Data Mining Software Churn Modeling Acquisition and Cross Sell Modeling


Relationship Marketing: Relationship Marketing Relationship Marketing is a Process communicating with your customers listening to their responses Companies take actions marketing campaigns new products new channels new packaging


Relationship Marketing -- continued: Relationship Marketing -- continued Customers and prospects respond most common response is no response This results in a cycle data is generated opportunities to learn from the data and improve the process emerge


The Move Towards Relationship Management: The Move Towards Relationship Management E-commerce companies want to customize the user experience Supermarkets want to be infomediaries Credit card companies want to recommend good restaurants and hotels in new cities Phone companies want to know your friends and family Bottom line: Companies want to be in the business of serving customers rather than merely selling products


CRM is Revolutionary: CRM is Revolutionary Grocery stores have been in the business of stocking shelves Banks have been in the business of managing the spread between money borrowed and money lent Insurance companies have been in the business of managing loss ratios Telecoms have been in the business of completing telephone calls Key point: More companies are beginning to view customers as their primary asset


Why Now ?: Why Now ? Representative Growth in a Maturing Market


The Electronic Trail: The Electronic Trail A customer places a catalog order over the telephone At the local telephone company time of call, number dialed, long distance company used, … At the long distance company (for the toll-free number) duration of call, route through switching system, … At the catalog items ordered, call center, promotion response, credit card used, inventory update, shipping method requested, …


The Electronic Trail-- continued: The Electronic Trail-- continued At the credit card clearing house transaction date, amount charged, approval code, vendor number, … At the bank billing record, interest rate, available credit update, … At the package carrier zip code, time stamp at truck, time stamp at sorting center, … Bottom line: Companies do keep track of data


An Illustration: An Illustration A few years ago, UPS went on strike FedEx saw its volume increase After the strike, its volume fell FedEx identified those customers whose FedEx volumes had increased and then decreased These customers were using UPS again FedEx made special offers to these customers to get all of their business


The Corporate Memory: The Corporate Memory Several years ago, Land’s End could not recognize regular Christmas shoppers some people generally don’t shop from catalogs but spend hundreds of dollars every Christmas if you only store 6 months of history, you will miss them Victoria’s Secret builds customer loyalty with a no-hassle returns policy some “loyal customers” return several expensive outfits each month they are really “loyal renters”


CRM Requires Learning and More: CRM Requires Learning and More Form a learning relationship with your customers Notice their needs On-line Transaction Processing Systems Remember their preferences Decision Support Data Warehouse Learn how to serve them better Data Mining Act to make customers more profitable


The Importance of Channels: The Importance of Channels Channels are the way a company interfaces with its customers Examples Direct mail Email Banner ads Telemarketing Billing inserts Customer service centers Messages on receipts Key data about customers come from channels


Channels -- continued: Channels -- continued Channels are the source of data Channels are the interface to customers Channels enable a company to get a particular message to a particular customer Channel management is a challenge in organizations CRM is about serving customers through all channels


Where Does Data Mining Fit In?: Where Does Data Mining Fit In? Hindsight Foresight Insight Analysis and Reporting (OLAP) Statistical Modeling Data Mining


Our Definition of Data Mining: Our Definition of Data Mining Exploration and analysis of large quantities of data By automatic or semi-automatic means To discover meaningful patterns and rules These patterns allow a company to better understand its customers improve its marketing, sales, and customer support operations Source: Berry and Linoff (1997)


Data Mining for Insight: Data Mining for Insight Classification Prediction Estimation Automatic Cluster Detection Affinity Grouping Description


Finding Prospects: Finding Prospects A cellular phone company wanted to introduce a new service They wanted to know which customers were the most likely prospects Data mining identified “sphere of influence” as a key indicator of likely prospects Sphere of influence is the number of different telephone numbers that someone calls


Paying Claims: Paying Claims A major manufacturer of diesel engines must also service engines under warranty Warranty claims come in from all around the world Data mining is used to determine rules for routing claims some are automatically approved others require further research Result: The manufacturer saves millions of dollars Data mining also enables insurance companies and the Fed. Government to save millions of dollars by not paying fraudulent medical insurance claims


Cross Selling: Cross Selling Cross selling is another major application of data mining What is the best additional or best next offer (BNO) to make to each customer? E.g., a bank wants to be able to sell you automobile insurance when you get a car loan The bank may decide to acquire a full-service insurance agency


Holding on to Good Customers: Holding on to Good Customers Berry and Linoff used data mining to help a major cellular company figure out who is at risk for attrition And why are they at risk They built predictive models to generate call lists for telemarketing The result was a better focused, more effective retention campaign


Weeding out Bad Customers: Weeding out Bad Customers Default and personal bankruptcy cost lenders millions of dollars Figuring out who are your worst customers can be just as important as figuring out who are your best customers many businesses lose money on most of their customers


They Sometimes get Their Man: They Sometimes get Their Man The FBI handles numerous, complex cases such as the Unabomber case Leads come in from all over the country The FBI and other law enforcement agencies sift through thousands of reports from field agents looking for some connection Data mining plays a key role in FBI forensics


Anticipating Customer Needs: Clustering is an undirected data mining technique that finds groups of similar items Based on previous purchase patterns, customers are placed into groups Customers in each group are assumed to have an affinity for the same types of products New product recommendations can be generated automatically based on new purchases made by the group This is sometimes called collaborative filtering Anticipating Customer Needs


CRM Focuses on the Customer: CRM Focuses on the Customer The enterprise has a unified view of each customer across all business units and across all channels This is a major systems integration task The customer has a unified view of the enterprise for all products and regardless of channel This requires harmonizing all the channels


A Continuum of Customer Relationships: A Continuum of Customer Relationships Large accounts have sales managers and account teams E.g., Coca-Cola, Disney, and McDonalds CRM tends to focus on the smaller customer --the consumer But, small businesses are also good candidates for CRM


What is a Customer: What is a Customer A transaction? An account? An individual? A household? The customer as a transaction purchases made with cash are anonymous most Web surfing is anonymous we, therefore, know little about the consumer


A Customer is an Account: A Customer is an Account More often, a customer is an account Retail banking checking account, mortgage, auto loan, … Telecommunications long distance, local, ISP, mobile, … Insurance auto policy, homeowners, life insurance, … Utilities The account-level view of a customer also misses the boat since each customer can have multiple accounts


Customers Play Different Roles: Customers Play Different Roles Parents buy back-to-school clothes for teenage children children decide what to purchase parents pay for the clothes parents “own” the transaction Parents give college-age children cellular phones or credit cards parents may make the purchase decision children use the product It is not always easy to identify the customer


The Customer’s Lifecycle: The Customer’s Lifecycle Childhood birth, school, graduation, … Young Adulthood choose career, move away from parents, … Family Life marriage, buy house, children, divorce, … Retirement sell home, travel, hobbies, … Much marketing effort is directed at each stage of life


The Customer’s Lifecycle is Unpredictable: The Customer’s Lifecycle is Unpredictable It is difficult to identify the appropriate events graduation, retirement may be easy marriage, parenthood are not so easy many events are “one-time” Companies miss or lose track of valuable information a man moves a woman gets married, changes her last name, and merges her accounts with spouse It is hard to track your customers so closely, but, to the extent that you can, many marketing opportunities arise


Customers Evolve Over Time: Customers Evolve Over Time Customers begin as prospects Prospects indicate interest fill out credit card applications apply for insurance visit your website They become new customers After repeated purchases or usage, they become established customers Eventually, they become former customers either voluntarily or involuntarily


Slide33: Business Processes Organize Around the Customer Lifecycle Acquisition Activation Relationship Management Winback Former Customer Prospect Established Customer New Customer Low Value High Potential High Value Voluntary Churn Forced Churn


Different Events OccurThroughout the Lifecycle: Different Events Occur Throughout the Lifecycle Prospects receive marketing messages When they respond, they become new customers They make initial purchases They become established customers and are targeted by cross-sell and up-sell campaigns Some customers are forced to leave (cancel) Some leave (cancel) voluntarily Others simply stop using the product (e.g., credit card) Winback/collection campaigns


Different Data is AvailableThroughout the Lifecycle: Different Data is Available Throughout the Lifecycle The purpose of data warehousing is to keep this data around for decision-support purposes Charles Schwab wants to handle all of their customers’ investment dollars Schwab observed that customers started with small investments


Different Data is AvailableThroughout the Lifecycle -- continued: Different Data is Available Throughout the Lifecycle -- continued By reviewing the history of many customers, Schwab discovered that customers who transferred large amounts into their Schwab accounts did so soon after joining After a few months, the marketing cost could not be justified Schwab’s marketing strategy changed as a result


Different Models are Appropriateat Different Stages: Different Models are Appropriate at Different Stages Prospect acquisition Prospect product propensity Best next offer Forced churn Voluntary churn Bottom line: We use data mining to predict certain events during the customer lifecycle


Different Approaches to Data Mining: Different Approaches to Data Mining Outsourcing let an outside expert do the work have him/her report the results Off-the-shelf, turn-key software solutions packages have generic churn models & response models they work pretty well Master Data Mining develop expertise in-house use sophisticated software such as Clementine or Enterprise Miner


Privacy is a Serious Matter: Privacy is a Serious Matter Data mining and CRM raise some privacy concerns These concerns relate to the collection of data, more than the analysis of data The next few slides illustrate marketing mistakes that can result from the abundance and availability of data


Using Data Mining to Help Diabetics: Using Data Mining to Help Diabetics Early detection of diabetes can save money by preventing more serious complications Early detection of complications can prevent worsening retinal eye exams every 6 or 12 months can prevent blindness these eye exams are relatively inexpensive So one HMO took action they decided to encourage their members, who had diabetes to get eye exams the IT group was asked for a list of members with diabetes


One Woman’s Response: One Woman’s Response Letters were sent out to HMO members Three types of diabetes – congenital, adult-onset, gestational One woman contacted had gestational diabetes several years earlier She was “traumatized” by the letter, thinking the diabetes had recurred She threatened to sue the HMO Mistake: Disconnect between the domain expertise and data expertise


Gays in the Military: Gays in the Military The “don’t ask; don’t tell” policy allows discrimination against openly gay men and lesbians in the military Identification as gay or lesbian is sufficient grounds for discharge This policy is enforced Approximately 1000 involuntary discharges each year


The Story of Former Senior ChiefPetty Officer Timothy McVeigh: The Story of Former Senior Chief Petty Officer Timothy McVeigh Several years ago, McVeigh used an AOL account, with an anonymous alias Under marital status, he listed “gay” A colleague discovered the account and called AOL to verify that the owner was McVeigh AOL gave out the information over the phone McVeigh was discharged (three years short of his pension) The story doesn’t end here


Two Serious Privacy Violations: Two Serious Privacy Violations AOL breached its own policy by giving out confidential user information AOL paid an undisclosed sum to settle with McVeigh and suffered bad press as well The law requires that government agents identify themselves to get online subscription information This was not done McVeigh received an honorable discharge with full retirement pension


Friends, Family, and Others: Friends, Family, and Others In the 1990s, MCI promoted the “Friends and Family” program They asked existing customers for names of people they talked with often If these friends and family signed up with MCI, then calls to them would be discounted Did MCI have to ask customers about who they call regularly? Early in 1999, BT (formerly British Telecom) took the idea one step beyond BT invented a new marketing program discounts to the most frequently called numbers


BT Marketing Program: BT Marketing Program BT notified prospective customers of this program by sending them their most frequently called numbers One woman received the letter uncovered her husband’s cheating threw him out of the house sued for divorce The husband threatened to sue BT for violating his privacy BT suffered negative publicity


No Substitute for Human Intelligence: No Substitute for Human Intelligence Data mining is a tool to achieve goals The goal is better service to customers Only people know what to predict Only people can make sense of rules Only people can make sense of visualizations Only people know what is reasonable, legal, tasteful Human decision makers are critical to the data mining process


A Long, Long Time Ago: A Long, Long Time Ago There was no marketing There were few manufactured goods Distribution systems were slow and uncertain There was no credit Most people made what they needed at home There were no cell phones There was no data mining It was sufficient to build a quality product and get it to market


Then and Now: Then and Now Before supermarkets, a typical grocery store carried 800 different items A typical grocery store today carries tens of thousands of different items There is intense competition for shelf space and premium shelf space In general, there has been an explosion in the number of products in the last 50 years Now, we need to anticipate and create demand (e.g., e-commerce) This is what marketing is all about


Effective Marketing Presupposes: Effective Marketing Presupposes High quality goods and services Effective distribution of goods and services Adequate customer service Marketing promises are kept Competition direct (same product) “wallet-share” Ability to interact directly with customers


The ACME Corporation: The ACME Corporation Imagine a fictitious corporation that builds widgets It can sell directly to customers via a catalog or the Web maintain control over brand and image It can sell directly through retail channels get help with marketing and advertising It can sell through resellers outsource marketing and advertising entirely Let’s assume ACME takes the direct marketing approach


Before Focusing on One-to-OneMarketing: Before Focusing on One-to-One Marketing Branding is very important provides a mark of quality to consumers old concept – Bordeaux wines, Chinese porcelain, Bruges cloth really took off in the 20th Century Advertising is hard media mix problem – print, radio, TV, billboard, Web difficult to measure effectiveness “Half of my advertising budget is wasted; I just don’t know which half.”


Different Approaches to Direct Marketing: Different Approaches to Direct Marketing Naïve Approach get a list of potential customers send out a large number of messages and repeat Wave Approach send out a large number of messages and test Staged Approach send a series of messages over time Controlled Approach send out messages over time to control response (e.g., get 10,000 responses/week)


The World is Speeding Up: The World is Speeding Up Advertising campaigns take months market research design and print material Catalogs are planned seasons in advance Direct mail campaigns also take months Telemarketing campaigns take weeks Web campaigns take days modification/refocusing is easy


How Data Mining Helps inMarketing Campaigns: How Data Mining Helps in Marketing Campaigns Improves profit by limiting campaign to most likely responders Reduces costs by excluding individuals least likely to respond AARP mails an invitation to those who turn 50 they excluded the bottom 10% of their list response rate did not suffer


How Data Mining Helps inMarketing Campaigns--continued: How Data Mining Helps in Marketing Campaigns--continued Predicts response rates to help staff call centers, with inventory control, etc. Identifies most important channel for each customer Discovers patterns in customer data


Some Background on ACME: Some Background on ACME They are going to pursue a direct marketing approach Direct mail marketing budget is $300,000 Best estimates indicate between 1 and 10 million customers ACME wants to target the customer base cost-effectively ACME seeks to assign a “score” to each customer which reflects the relative likelihood of that customer purchasing the product


How Do You Assign Scores: How Do You Assign Scores Randomly – everyone gets the same score Assign relative scores based on ad-hoc business knowledge Assign a score to each cell in an RFM (recency, frequency, monetary) analysis Profile existing customers and use these profiles to assign scores to similar, potential customers Build a predictive model based on similar product sales in the past


Data Mining Models Assign a Scoreto Each Customer: ID Name State Score Rank 0102 Will MA 0.314 7 0104 Sue NY 0.159 9 0105 John AZ 0.265 8 0110 Lori AZ 0.358 5 0111 Beth NM 0.979 1 0112 Pat WY 0.328 6 0116 David ID 0.446 4 0117 Frank MS 0.897 2 0118 Ethel NE 0.446 4 Data Mining Models Assign a Score to Each Customer Comments 1. Think of score as likelihood of responding 2. Some scores may be the same


Approach 1: Budget Optimization: Approach 1: Budget Optimization ACME has a budget of $300,000 for a direct mail campaign Assumptions each item being mailed costs $1 this cost assumes a minimum order of 20,000 ACME can afford to contact 300,000 customers ACME contacts the highest scoring 300,000 customers Let’s assume ACME is selecting from the top three deciles


The Concept of Lift: The Concept of Lift If we look at a random 10% of the potential customers, we expect to get 10% of likely responders Can we select 10% of the potential customers and get more than 10% of likely responders? If so, we realize “lift” This is a key goal in data mining


Slide63: Notes x-axis gives population percentile y-axis gives the lift the top 10% of the scorers are 3 times more likely to respond than a random 10% would be The Actual Lift Chart


How Well Does ACME Do?: How Well Does ACME Do? ACME selects customers from the top three deciles From cumulative gains chart, a response rate of 65% (vs. 30%) results From lift chart, we see a lift of 65/30 = 2.17 The two charts convey the same information, but in different ways


Can ACME Do Better?: Can ACME Do Better? Test marketing campaign send a mailing to a subset of the customers, say 30,000 take note of the 1 to 2% of those who respond build predictive models to predict response use the results from these models The key is to learn from the test marketing campaign


Optimizing the Budget: Optimizing the Budget Decide on the budget Based on cost figures, determine the size of the mailing Develop a model to score all customers with respect to their relative likelihood to respond to the offer Choose the appropriate number of top scoring customers


Approach 2: Optimizing the Campaign: Approach 2: Optimizing the Campaign Lift allows us to contact more of the potential responders It is a very useful measure But, how much better off are we financially? We seek a profit-and-loss statement for the campaign To do this, we need more information than before


Is the Campaign Profitable?: Is the Campaign Profitable? Suppose the following the typical customer will purchase about $100 worth of merchandise from the next catalog of the $100, $55 covers the cost of inventory, warehousing, shipping, and so on the cost of sending mail to each customer is $1 Then, the net revenue per customer in the campaign is $100 - $55 - $1 = $44


The Profit/Loss Matrix: Someone who scores in the top 30%, is predicted to respond Those predicted to respond cost $1 those who actually respond yield a gain of $45 those who don’t respond yield no gain Those not predicted to respond cost $0 and yield no gain The Profit/Loss Matrix ACTUAL Predicted


The Profit/Loss Matrix--continued: The Profit/Loss Matrix--continued The profit/loss matrix is a powerful concept But, it has its limitations people who don’t respond become more aware of the brand/product due to the marketing campaign they may respond next time people not contacted might have responded had they been invited For now, let’s focus on the profit/loss matrix


How Do We Get the P/L Numbers?: How Do We Get the P/L Numbers? Cost numbers are relatively easy mailing and printing costs can be handled by accounts payable call center costs, for incoming orders, are usually fixed Revenue numbers are rough estimates based on previous experience, back-of-envelope calculations, guesswork based on models of customer buying behavior


Is the Campaign Profitable?: Is the Campaign Profitable? Assumptions made so far $44 net revenue per responder ($1) net revenue per non-responder 300,000 in target group new assumption: overhead charge of $20,000 Resulting lift is 2.17 We can now estimate profit for different response rates


Net Revenue for the Campaign: Net Revenue for the Campaign The campaign makes money if it achieves a response rate of at least 3%


Net Revenue Table Explained: Suppose response rate of 3% Net revenue = 9000 44 + 291,000 (-1) - 20,000 = $85,000 Lift = response rate for campaign overall response rate overall response rate = response rate for campaign lift = = 1.38% Suppose response rate of 6% Net Revenue Table Explained


Two Ways to Estimate Response Rates: Two Ways to Estimate Response Rates Use a randomly selected hold-out set (the test set) this data is not used to build the model the model’s performance on this set estimates the performance on unseen data Use a hold-out set on oversampled data most data mining involves binary outcomes often, we try to predict a rare event (e.g., fraud) with oversampling, we overrepresent the rare outcomes and underrepresent the common outcomes


Oversampling Builds Better Models for Rare Events: Oversampling Builds Better Models for Rare Events Suppose 99% of records involve no fraud A model that always predicts no fraud will be hard to beat But, such a model is not useful Stratified sampling with two outcomes is called oversampling


Return to Earlier Model: Return to Earlier Model


Assume an Overall Response Rate of 1%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1% and Calculate the Profit for Each Decile Remember: $44 net revenue/ $1 cost per item mailed/ $20,000 overhead


Review of Profit Calculation: Key equations size (yes) = lift profit = 44 size (yes) - size (no) - 20,000 Example: top three deciles (30% row) size (yes) = 2.167 = 6500 profit = 286,000 - 293,500 - 20,000 = -27,500 Notice that top 10% yields the maximum profit Mailing to the top three deciles would cost us money Review of Profit Calculation size


Typical Shape for a Profit Curve($44, $1, $20,000): Typical Shape for a Profit Curve ($44, $1, $20,000)


Approach 2 Summary: Approach 2 Summary Estimate cost per contact, overhead, and estimated revenue per responder Build a model and estimate response probabilities for each customer Order the customers by their response scores For each decile, calculate the cumulative number of responders and non-responders Using the estimates, determine the cumulative profit for each decile Choose all the deciles up to the one with the highest cumulative profit


The Problem with Campaign Optimization: The Problem with Campaign Optimization Campaign optimization is very sensitive to the underlying assumptions Suppose the response rate is 2% rather than 1%? Suppose the cost of contacting a customer is $1.20 rather than $1? Sensitivity is a serious problem


Assume an Overall Response Rate of 1.2%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1.2% and Calculate the Profit for Each Decile Remember: $44 net revenue/ $1 cost per item mailed/ $20,000 overhead


Assume an Overall Response Rate of 0.8%and Calculate the Profit for Each Decile: Remember: $44 net revenue/ $1 cost per item mailed/ $20,000 overhead Assume an Overall Response Rate of 0.8% and Calculate the Profit for Each Decile


Assume an Overall Response Rate of 2%and Calculate the Profit for Each Decile: Remember: $44 net revenue/ $1 cost per item mailed/ $20,000 overhead Assume an Overall Response Rate of 2% and Calculate the Profit for Each Decile


Dependence on Response Rate($44, $1, $20,000): Dependence on Response Rate ($44, $1, $20,000)


Assume an Overall Response Rate of 1%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1% and Calculate the Profit for Each Decile Remember: $44 net revenue/ $1.2 cost per item mailed/ $20,000 overhead


Assume an Overall Response Rate of 1%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1% and Calculate the Profit for Each Decile Remember: $44 net revenue/ $0.8 cost per item mailed/ $20,000 overhead


Assume an Overall Response Rate of 1%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1% and Calculate the Profit for Each Decile Remember: $44 net revenue/ $2 cost per item mailed/ $20,000 overhead


Dependence on Costs: Dependence on Costs


Assume an Overall Response Rate of 1%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1% and Calculate the Profit for Each Decile Remember: $35.2 net revenue/ $1 cost per item mailed/ $20,000 overhead


Assume an Overall Response Rate of 1%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1% and Calculate the Profit for Each Decile Remember: $52.8 net revenue/ $1 cost per item mailed/ $20,000 overhead


Assume an Overall Response Rate of 1%and Calculate the Profit for Each Decile: Assume an Overall Response Rate of 1% and Calculate the Profit for Each Decile Remember: $88 net revenue/ $1 cost per item mailed/ $20,000 overhead


Dependence on Revenue: Dependence on Revenue


Campaign Optimization Drawbacks: Campaign Optimization Drawbacks Profitability depends on response rates, cost estimates, and revenue potential Each one impacts profitability The numbers we use are just estimates If we are off by a little here and a little there, our profit estimates could be off by a lot In addition, the same group of customers is chosen for multiple campaigns


Approach 3: Customer Optimization: Approach 3: Customer Optimization Campaign optimization makes a lot of sense But, campaign profitability is difficult to estimate Is there a better way? Do what is best for each customer Focus on customers, rather than campaigns


Real-World Campaigns: Real-World Campaigns Companies usually have several products that they want to sell telecom: local, long distance, mobile, ISP, etc. banking: CDs, mortgages, credit cards, etc. insurance: home, car, personal liability, etc. retail: different product lines There are also upsell and customer retention programs These campaigns compete for customers


Each Campaign May Have a Separate Model: Each Campaign May Have a Separate Model These models produce scores The score tells us how likely a given customer is to respond to that specific campaign 0, if the customer already has the product 0, if the product and customer are incompatible 1, if the customer has asked about the product Each campaign is relevant for a subset of all the customers Imagine three marketing campaigns, each with a separate data mining model


Sample Scores (as Rankings for Three Different Campaigns): Sample Scores (as Rankings for Three Different Campaigns) ID Name State Mod A Mod B Mod C 0102 Will MA 3 4 2 0104 Sue NY 1 2 4 0105 John AZ 2 1 1 0110 Lori AZ 5 7 6 0111 Beth NM 9 3 8 0112 Pat WY 4 5 2 0116 David ID 6 5 7 0117 Frank MS 8 9 8 0118 Ethel NE 6 8 5


Choose the Best Customers for Each Campaign: Choose the Best Customers for Each Campaign


A Common Situation: A Common Situation “Good” customers are typically targeted by many campaigns Many other customers are not chosen for any campaigns Good customers who become inundated with contacts become less likely to respond at all Let the campaigns compete for customers


Choose the Best Campaign for Each Customer: Choose the Best Campaign for Each Customer


Focus on the Customer: Focus on the Customer Determine the propensity of each customer to respond to each campaign Estimate the net revenue for each customer from each campaign Incorporate profitability into the customer-optimization strategy Not all campaigns will apply to all customers


First, Determine Response Rate for Each Campaign: First, Determine Response Rate for Each Campaign Customers who are not candidates are given a rate of zero


Second, Add in Product Profitability: Second, Add in Product Profitability As a more sophisticated alternative, profit could be estimated for each customer/product combination


Finally, Determine the Campaign with the Highest Value: Finally, Determine the Campaign with the Highest Value EP (k) = the expected profit of product k For each customer, choose the highest expected profit campaign


Conflict Resolution with Multiple Campaigns: Conflict Resolution with Multiple Campaigns Managing many campaigns at the same time is complex for technical and political reasons Who owns the customer? Handling constraints each campaign is appropriate for a subset of customers each campaign has a minimum and maximum number of contacts each campaign seeks a target response rate new campaigns emerge over time


Marketing Campaigns and CRM: Marketing Campaigns and CRM The simplest approach is to optimize the budget using the rankings that models produce Campaign optimization determines the most profitable subset of customers for a given campaign, but it is sensitive to assumptions Customer optimization is more sophisticated It chooses the most profitable campaign for each customer


The Data Mining Process: The Data Mining Process What role does data mining play within an organization? How does one do data mining correctly? The SEMMA Process select and sample explore modify model assess


Identify the Right Business Problem: Identify the Right Business Problem Involve the business users Have them provide business expertise, not technical expertise Define the problem clearly “predict the likelihood of churn in the next month for our 10% most valuable customers” Define the solution clearly is this a one-time job, an on-going monthly batch job, or a real-time response (call centers and web)? What would the ideal result look like? how would it be used?


Transforming the Data into Actionable Information: Transforming the Data into Actionable Information Select and sample by extracting a portion of a large data set-- big enough to contain significant information, but small enough to manipulate quickly Explore by searching for unanticipated trends and anomalies in order to gain understanding


Transforming the Data into Actionable Information-- continued: Modify by creating, selecting, and transforming the variables to focus the model selection process Model by allowing the software to search automatically for a combination of variables that reliably predicts a desired outcome Assess by evaluating the usefulness and reliability of the findings from the data mining process Transforming the Data into Actionable Information-- continued


Act on Results: Act on Results Marketing/retention campaign lists or scores Personalized messages Customized user experience Customer prioritization Increased understanding of customers, products, messages


Measure the Results: Measure the Results Confusion matrix Cumulative gains chart Lift chart Estimated profit


Data Mining Uses Data from the Past to Effect Future Action: Data Mining Uses Data from the Past to Effect Future Action “Those who do not remember the past are condemned to repeat it.” – George Santayana Analyze available data (from the past) Discover patterns, facts, and associations Apply this knowledge to future actions


Examples: Examples Prediction uses data from the past to make predictions about future events (“likelihoods” and “probabilities”) Profiling characterizes past events and assumes that the future is similar to the past (“similarities”) Description and visualization find patterns in past data and assume that the future is similar to the past


We Want a Stable Model: We Want a Stable Model A stable model works (nearly) as well on unseen data as on the data used to build it Stability is more important than raw performance for most applications we want a car that performs well on real roads, not just on test tracks Stability is a constant challenge


Is the Past Relevant?: Is the Past Relevant? Does past data contain the important business drivers? e.g., demographic data Is the business environment from the past relevant to the future? in the ecommerce era, what we know about the past may not be relevant to tomorrow users of the web have changed since late 1990s Are the data mining models created from past data relevant to the future? have critical assumptions changed?


Data Mining is about Creating Models: Data Mining is about Creating Models A model takes a number of inputs, which often come from databases, and it produces one or more outputs Sometimes, the purpose is to build the best model The best model yields the most accurate output Such a model may be viewed as a black box Sometimes, the purpose is to better understand what is happening This model is more like a gray box


Models: Models Past Present Future Data ends here Actions take place here Building models takes place in the present using data from the past outcomes are already known Applying (or scoring) models takes place in the present Acting on the results takes place in the future outcomes are not known


Often, the Purpose is to Assign a Scoreto Each Customer: Often, the Purpose is to Assign a Score to Each Customer Comments Scores are assigned to rows using models Some scores may be the same 3. The scores may represent the probability of some outcome


Common Examples of What a Score Could Mean: Common Examples of What a Score Could Mean Likelihood to respond to an offer Which product to offer next Estimate of customer lifetime Likelihood of voluntary churn Likelihood of forced churn Which segment a customer belongs to Similarity to some customer profile Which channel is the best way to reach the customer


The Scores Provide a Rankingof the Customers: The Scores Provide a Ranking of the Customers SORT


This Ranking give Rise to Quantiles(terciles, quintiles, deciles, etc.): This Ranking give Rise to Quantiles (terciles, quintiles, deciles, etc.) } } } high medium low


Layers of Data Abstraction: Layers of Data Abstraction SEMMA starts with data There are many different levels of data within an organization Think of a pyramid The most abundant source is operational data every transaction, bill, payment, etc. at bottom of pyramid Business rules tell us what we’ve learned from the data at top of pyramid Other layers in between


SEMMA: Select and Sample: SEMMA: Select and Sample What data is available? Where does it come from? How often is it updated? When is it available? How recent is it? Is internal data sufficient? How much history is needed?


Data Mining Prefers Customer Signatures: Data Mining Prefers Customer Signatures Often, the data come from many different sources Relational database technology allows us to construct a customer signature from these multiple sources The customer signature includes all the columns that describe a particular customer the primary key is a customer id the target columns contain the data we want to know more about (e.g., predict) the other columns are input columns


Profiling is a Powerful Tool: Profiling is a Powerful Tool Profiling involves finding patterns from the past and assuming they will remain valid The most common approach is via surveys Surveys tell us what our customers and prospects look like Typical profiling question: What do churners look like? Profiling is frequently based on demographic variables e.g., location, gender, age


Profiling has its Limitations: Profiling has its Limitations Even at its best, profiling tells us about the past Connection between cause and effect is sometimes unclear people with brokerage accounts have a minimal balance in their savings account customers who churn are those who have not used their telephones (credit cards) for the past month customers who use voicemail make a lot of short calls to the same number More appropriate for advertising than one-to-one marketing


Two Ways to Aim for the Target: Two Ways to Aim for the Target Profiling: What do churners look like? data in input columns can be from the same time period (the past) as the target Prediction: Build a model that predicts who will churn next month data from input columns must happen before the target data comes from the past the present is when new data are scored


The Past Needs to Mimic the Present: The Past Needs to Mimic the Present Past Present Future Distant Past ends here Recent Past starts here Data ends here Predictions start here We mimic the present by using the distant past to predict the recent past


How Data from Different Time Periods are Used: How Data from Different Time Periods are Used Jan Feb Mar Apr May Jun Jul Aug Sep Model Set Score Set The model set is used to build the model The score set is used to make predictions It is now August X marks the month of latency Numbers to left of X are months in the past


Multiple Time Windows Help the ModelsDo Well in Predicting the Future: Multiple Time Windows Help the Models Do Well in Predicting the Future Jan Feb Mar Apr May Jun Jul Aug Sep Model Set Score Set Multiple time windows capture a wider variety of past behavior They prevent us from memorizing a particular season


Rules for Building a Model Set fora Prediction: Rules for Building a Model Set for a Prediction All input columns must come strictly before the target There should be a period of “latency” corresponding to the time needed to gather the data The model set should contain multiple time windows of data


More about the Model and Score Sets: More about the Model and Score Sets The model set can be partitioned into three subsets the model is trained using pre-classified data called the training set the model is refined, in order to prevent memorization, using the test set the performance of models can be compared using a third subset called the evaluation or validation set The model is applied to the score set to predict the (unknown) future


Stability Challenge: Memorizingthe Training Set: Stability Challenge: Memorizing the Training Set Training Data Error Rate Model Complexity Decision trees and neural networks can memorize nearly any pattern in the training set


Danger: Overfitting: Danger: Overfitting Danger: Overfitting The model has overfit the training data As model complexity grows, performance deteriorates on test data Model Complexity Training Data Test Data This is the model we want Error Rate


Building the Model from Data: Building the Model from Data Both the training set and the test set are used to create the model Algorithms find all the patterns in the training set some patterns are global (should be true on unseen data) some patterns are local (only found in the training set) We use the test set to distinguish between the global patterns and the local patterns Finally, the validation set is needed to evaluate the model’s performance


SEMMA: Explore the Data: SEMMA: Explore the Data Look at the range and distribution of all the variables Identify outliers and most common values Use histograms, scatter plots, and subsets Use algorithms such as clustering and market basket analysis Clementine does some of this for you when you load the data


SEMMA: Modify: SEMMA: Modify Add derived variables total, percentages, normalized ranges, and so on extract features from strings and codes Add derived summary variables median income in ZIP code Remove unique, highly skewed, and correlated variables often replacing them with derived variables Modify the model set


The Density Problem: The Density Problem The model set contains a target variable “fraud” vs. “not fraud” “churn” vs. “still a customer” Often binary, but not always The density is the proportion of records with the given property (often quite low) fraud ≈ 1% churn ≈ 5% Predicting the common outcome is accurate, but not helpful


Back to Oversampling: Back to Oversampling Original data has 45 white and 5 dark (10% density) The model set has 10 white and 5 dark (33% density ) For every 9 white (majority) records in the original data, two are in the oversampled model set Oversampling rate is 9/2 = 4.5 1 11 21 31 41 2 12 22 32 42 3 13 23 33 43 4 14 24 34 44 5 15 25 35 45 6 16 26 36 46 7 17 27 37 47 8 18 28 37 48 9 19 29 39 49 10 20 30 40 50 10 20 30 40 50 2 6 12 17 23 29 31 34 45 48


Two Approaches to Oversampling: Two Approaches to Oversampling Build a new model set of the desired density fewer rows takes less time to build models more time for experimentation in practice, aim for at least 10,000 rows Use frequencies to reduce the importance of some rows uses all of the data Use a density of approx. 50% for binary outcomes


Oversampling by Taking a Subset ofthe Model Set: Oversampling by Taking a Subset of the Model Set The original data has 2 Ts and 7 Fs (22% density) Take all the Ts and 4 of the Fs (33% density) The oversampling rate is 7/4 = 1.75


Oversampling via Frequencies: Oversampling via Frequencies Add a frequency or weight column for each F, Frq = 0.5 for each T, Frq = 1.0 The model set has density of 2/(2 + 0.5 7) = 36.4% The oversampling rate is 7/3.5 = 2


SEMMA: Model: SEMMA: Model Choose an appropriate technique decision trees neural networks regression combination of above Set parameters Combine models


Regression: Regression Tries to fit data points to a known curve (often a straight line) Standard (well-understood) statistical technique Not a universal approximator (form of the regression needs to be specified in advance)


Neural Networks: Neural Networks Based loosely on computer models of how brains work Consist of neurons (nodes) and arcs, linked together Each neuron applies a nonlinear function to its inputs to produce an output Particularly good at producing numeric outputs No explanation of result is provided


Decision Trees: Decision Trees Looks like a game of “Twenty Questions” At each node, we fork based on variables e.g., is household income less than $40,000? These nodes and forks form a tree Decision trees are useful for classification problems especially with two outcomes Decision trees explain their result the most important variables are revealed


Experiment to Find the Best Modelfor Your Data: Experiment to Find the Best Model for Your Data Try different modeling techniques Try oversampling at different rates Tweak the parameters Add derived variables Remember to focus on the business problem


It is Often Worthwhile to Combinethe Results from Multiple Models: It is Often Worthwhile to Combine the Results from Multiple Models


Multiple-Model Voting: Multiple-Model Voting Multiple models are built using the same input data Then a vote, often a simple majority or plurality rules vote, is used for the final classification Requires that models be compatible Tends to be robust and can return better results


Segmented Input Models: Segmented Input Models Segment the input data by customer segment by recency Build a separate model for each segment Requires that model results be compatible Allows different models to focus and different models to use richer data


Combining Models: Combining Models What is response to a mailing from a non-profit raising money (1998 data set) Exploring the data revealed the more often, the less money one contributes each time so, best customers are not always most frequent Thus, two models were developed who will respond? how much will they give?


Compatible Model Results: Compatible Model Results In general, the score refers to a probability for decision trees, the score may be the actual density of a leaf node for a neural network, the score may be interpreted as the probability of an outcome However, the probability depends on the density of the model set The density of the model set depends on the oversampling rate


An Example: An Example The original data has 10% density The model set has 33% density Each white in model set represents 4.5 white in original data Each dark represents one dark The oversampling rate is 4.5


A Score Represents a Portion of the Model Set: A Score Represents a Portion of the Model Set Suppose an algorithm identifies the group at right as most likely to churn The score would be 4/6 = 67%, versus the density of 33% for the entire model set This score represents the probability on the oversampled data This group has a lift of 67/33 = 2


Determining the Score on the Original Data: Determining the Score on the Original Data The corresponding group in the original data has 4 dark and 9 white, for a score of 4 / (4 + 9) = 30.7% The original data has a density of 10% The lift is now 30.7/10 = 3.07


Determining the Score -- continued: Determining the Score -- continued The original group accounted for 6/15 = 40% of the model set In the original data, it corresponds to 13/50 = 26% Bottom line: before comparing the scores that different models produce, make sure that these scores are adjusted for the oversampling rate The final part of the SEMMA process is to assess the results


Confusion Matrix (or Correct Classification Matrix): Confusion Matrix (or Correct Classification Matrix) When the model predicts No, it is right 100/150 = 67% of the time The density of the model set is 150/1000 = 15% There are 1000 records in the model set When the model predicts Yes, it is right 800/850 = 94% of the time Yes No 800 50 50 100 Yes No Predicted Actual


Confusion Matrix-- continued: Confusion Matrix-- continued The model is correct 800 times in predicting Yes The model is correct 100 times in predicting No The model is wrong 100 times in total The overall prediction accuracy is 900/1000 = 90%


From Data to the Confusion Matrix: From Data to the Confusion Matrix T F 2 2 1 4 T F Predicted Actual We hold back a portion of the data so we have scores and actual values The top tercile is given a predicted value of T Because of tie, we have 4 Ts predicted


How Oversampling Affects the Results: How Oversampling Affects the Results The model set has a density of 15% No Suppose we achieve this density with an oversampling rate of 10 So, for every Yes in the model set there are 10 Yes’s in the original data Yes No 800 50 50 100 Yes No Predicted Actual Yes No 8000 50 500 100 Yes No Predicted Actual Model set Original data


How Oversampling Affects the Results--continued: How Oversampling Affects the Results--continued Original data has a density of 150/8650 = 1.734% We expect the model to predict No correctly 100/600 = 16.7% of the time The accuracy has gone down from 67% to 16.7% The results will vary based upon the degree of oversampling


Lift Measures How Well the Model is Doing: Lift Measures How Well the Model is Doing The lift is 66.7/33.3 = 2 The model is doing twice as well as choosing circles at random The density of dark in the model set is 33.3% The density of dark in the subset chosen by the model is 66.7%


Lift on a Small Data Set: Lift on a Small Data Set Tercile 1 has a density of 66.7% The lift is 66.7/33.3 = 2 Note: we break tie arbitrarily Model set density of T is 33.3% Tercile 1 has two T and one F


The Lift Chart for the Small Data Set: The Lift Chart for the Small Data Set We always look at lift in a cumulative sense 1 2 3 Tercile 1 has a lift of 66.7/33.3 = 2 Terciles 1 and 2 have a density of 3/6 = 50% and a lift of 50/33.3 = 1.5 Since terciles 1, 2, and 3 comprise the entire model set, the lift is 1 2.5 2.0 1.5 1.0 0.5 Tercile


Cumulative Gains Chart: Cumulative Gains Chart The cumulative gains chart and the lift chart are related 100% 67% 33% 33% 67% 100% with model random model Cumulative gains chart shows the proportion of responders (churners) in each tercile (decile) Horizontal axis shows the tercile (decile) Vertical axis gives the proportion of responders that model yields * * * * *


More on Lift and the Cumulative Gains Chart: More on Lift and the Cumulative Gains Chart The lift curve is the ratio of the cumulative gains of the model to the cumulative gains of the random model The cumulative gains chart has several advantages it always goes up and to the right can be used diagnostically Reporting the lift or cumulative gains on the training set is cheating Reporting the lift or cumulative gains without the oversampling rate is also cheating


Summary of Data Mining Process: Summary of Data Mining Process Cycle of Data Mining identify the right business problem transform data into actionable information act on the information measure the effect SEMMA select/sample data to create the model set explore the data modify data as necessary model to produce results assess effectiveness of the models


The Data in Data Mining: The Data in Data Mining Data comes in many forms internal and external sources Different sources of data have different peculiarities Data mining algorithms rely on a customer signature one row per customer multiple columns See Chapter 6 in Mastering Data Mining for details


Preventing Customer Attrition: Preventing Customer Attrition We use the noun churn as a synonym for attrition We use the verb churn as a synonym for leave Why study attrition? it is a well-defined problem it has a clear business value we know our customers and which ones are valuable we can rely on internal data the problem is well-suited to predictive modeling


When You Know Who is Likely to Leave,You Can …: When You Know Who is Likely to Leave, You Can … Focus on keeping high-value customers Focus on keeping high-potential customers Allow low-potential customers to leave, especially if they are costing money Don’t intervene in every case Topic should be called “managing customer attrition”


The Challenge of Defining Customer Value: The Challenge of Defining Customer Value We know how customers behaved in the past We know how similar customers behaved in the past Customers have control over revenues, but much less control over costs we may prefer to focus on net revenue rather than profit We can use all of this information to estimate customer worth These estimates make sense for the near future


Representative Growth in a Maturing Market: Representative Growth in a Maturing Market Why Maturing Industries Care About Customer Attrition As growth flattens out, customer attrition becomes more important In this region of rapid growth, building infrastructure is more important than CRM total customers new customers churners


Another View of Customer Attrition: Another View of Customer Attrition In a fast growth market, almost all customers are new customers In a maturing market, customer attrition and customer profitability become issues In a mature market, the amount of attrition almost equals the number of new customers


Why Attrition is Important: Why Attrition is Important When markets are growing rapidly, attrition is usually not important customer acquisition is more important Eventually, every new customer is simply replacing one who left Before this point, it is cheaper to prevent attrition than to spend money on customer acquisition One reason is that, as a market becomes saturated, acquisition costs go up


In Maturing Markets, Acquisition Response Rates Decrease and Costs Increase: In Maturing Markets, Acquisition Response Rates Decrease and Costs Increase Assumption: $1 per contact, $20 offer 220 120 70 53.3 45 40


Acquisition versus Retention: Acquisition versus Retention Assumption: $1 per contact, $20 offer 220 120 70 53.3 45 40


Acquisition versus Retention-- continued: Acquisition versus Retention-- continued As response rate drops, suppose we spend $140 to obtain a new customer Alternatively, we could spend $140 to retain an existing customer Assume the two customers have the same potential value Some possible options decrease the costs of the acquisition campaign implement a customer retention campaign combination of both


Retention and Acquisition are Different: Retention and Acquisition are Different Economics of acquisition campaigns customers are unlikely to purchase unless contacted/invited cost of acquisition is the campaign cost divided by the number acquired during the campaign Economics of retention campaigns customers are targeted for retention, but some would have remained customers anyway cost of retention is the campaign cost divided by the net increase in retained customers


Cost per Retention: How ManyResponders Would Have Left?: Cost per Retention: How Many Responders Would Have Left? For a retention campaign, we need to distinguish between customers who merely respond and those who respond and would have left If the overall group has an attrition rate of 25%, you can assume that one quarter of the responders are saved Since people like to receive something for nothing, response rates tend to be high As before, assume $1 per contact and $20 per response We need to specify response rate and attrition rate


Slide183: Cost Per Retention by Attrition Rate


A Sample Calculation: A Sample Calculation Given: $1 per contact, $20 per response What is the cost per retention with response rate of 20% and attrition rate of 10%? Suppose 100 people are contacted 20 people respond 20 10% = 2 would have left, but are retained campaign cost = $100 + $20 20 = $500 cost per retention = $500/2 = $250


Typical Customer Retention Data: Typical Customer Retention Data Each year, 40% of existing customers leave Each year, 70% new customers are added At year end, the number of customers is the number at the end of the previous year minus the number who left plus the number of new customers


How to Lie with Statistics: How to Lie with Statistics When we divide by end-of-year customers, we reduce the attrition rate to about 31% per year This may look better, but it is cheating


Suppose Acquisition Suddenly Stops: Suppose Acquisition Suddenly Stops If acquisition of new customers stops, existing customers still leave at same rate But, churn rate more than doubles


Measuring Attrition the Right Wayis Difficult: Measuring Attrition the Right Way is Difficult The “right way” gives a value of 40% instead of 30.77% who wants to increase their attrition rate? this happens when the number of customers is increasing For our small example, we assume that new customers do not leave during their first year the real world is more complicated we might look at the number of customers on a monthly basis


Slide189: Assumption: $20 for existing customers; $10 for new customers Revenue Time Effect of Attrition on Revenue Effect of Attrition on Revenue


Effect of Attrition on Revenues -- continued: Effect of Attrition on Revenues -- continued From previous page, the loss due to attrition is about the same as the new customers revenue The loss due to attrition has a cumulative impact If we could retain some of these customers each year, they would generate revenue well into the future It is useful to examine the relationship between attrition and the length of the customer relationship Often, attrition is greater for longer-duration customers


Relationship Between Attrition and theLength of the Customer Relationship: Relationship Between Attrition and the Length of the Customer Relationship


What is a Customer Attrition Score?: What is a Customer Attrition Score? When building an attrition model, we seek a score for each customer This adds a new column to the data Two common approaches a relative score or ranking of who is going to leave an estimate of the likelihood of leaving in the next time period


Attrition Scores are an ImportantPart of the Solution: Attrition Scores are an Important Part of the Solution Having an idea about which customers will leave does not address key business issues why are they leaving? where are they going? is brand strength weakening? are the products still competitive? However, it does allow for more targeted marketing to customers it is often possible to gain understanding and combat attrition while using attrition models


Requirements of Effective Attrition Management: Requirements of Effective Attrition Management Keeping track of the attrition rate over time Understanding how different methods of acquisition impact attrition Looking at some measure of customer value, to determine which customers to let leave Implementing attrition retention efforts for high-value customers Knowing who might leave in the near future


Three Types of Attrition: Three Types of Attrition Voluntary attrition when the customer goes to a competitor our primary focus is on voluntary attrition models Forced attrition when the company decides that it no longer wants the customer and cancels the account often due to non-payment our secondary focus is on forced attrition models Expected attrition when the customer changes lifestyle and no longer needs your product or service


What is Attrition?: What is Attrition? In the telecommunications industry it is very clear customers pay each month for service customers must explicitly cancel service the customer relationship is primarily built around a single product It is not so clear in other industries retail banking credit cards retail e-commerce


Consider Retail Banking: Consider Retail Banking Customers may have a variety of accounts deposit and checking accounts savings and investment mortgages loans credit cards business accounts What defines attrition? closing one account? closing all accounts? something else?


Retail Banking--Continued: Retail Banking--Continued One large bank uses checking account balance to define churn What if a customer (e.g., credit card or mortgage customer) does not have a checking account with the bank? Another example from retail banking involves forced attrition Consider the bank loan timeline on the next page


The Path to Forced Attrition: The Path to Forced Attrition Forced Attrition: the bank sells the loan to a collection agency


Credit Cards: Credit Cards Credit cards, in the U.S., typically don’t have an annual fee there is no incentive for a cardholder to tell you when she no longer plans on using the card Cardholders typically use the card several times each month Customers who stop using the card and are not carrying a balance are considered to be silent churners The definition of silent churn varies among issuers customers who have not used the card for six months customers who have not used the card for three months and for nine of the last twelve months


The Catalog Industry: The Catalog Industry Like credit cards, catalogs are free therefore, customers have little incentive to cancel them Unlike credit cards, purchases from catalogs are more sporadic so defining silent churn is more difficult Purchases from catalogs are often seasonal so silent churn is measured over the course of years, rather than months


E-commerce: E-commerce Attrition in e-commerce is hardest of all to define consider Amazon Web sites are free, so there is no incentive to announce intention to churn Customers often visit many times before making a purchase Some customers buy frequently and others don’t how can Amazon decrease time between purchases? The e-commerce world is still growing rapidly, so customer retention is not a major issue But it will become a major issue soon


Ways to Address Voluntary Attrition: Ways to Address Voluntary Attrition Allow customers who are not valuable to leave Offer incentives to stay around for a period of time teaser rates on credit cards, free weekend airtime, no payments or interest for six months Offer incentives to valuable customers discounts/extras Stimulate usage miles for minutes, donations to charities, discounts


Predicting Voluntary Attrition Can Be Dangerous: Predicting Voluntary Attrition Can Be Dangerous Voluntary attrition sometimes looks like expected or forced attrition Confusing voluntary and expected attrition results in the waste of marketing dollars Confusing voluntary and forced attrition means you lose twice by, again, wasting marketing dollars by incurring customer non-payment


Ways to Address Forced Attrition: Ways to Address Forced Attrition Stop marketing to the customer no more catalogs, billing inserts, or other usage stimulation Reduce credit lines Increase minimum payments Accelerate cancellation timeline Increase interest rates to increase customer value while he is still paying


Predicting Forced Attrition Can Be Dangerous: Predicting Forced Attrition Can Be Dangerous Forced attrition sometimes looks at good customers they have high balances they are habitually late Confusing forced attrition with good customers may encourage them to leave Confusing forced with voluntary attrition may hamper winback prospects


Attrition Scores are Designed for theNear Future: Attrition Scores are Designed for the Near Future The chance that someone will leave tomorrow is essentially 0% The chance that someone will leave in the next 100 years is essentially 100% A good attrition score is valid for one to three months in the future Attrition modeling can take place at any point in the customer lifecycle


A Related Problem: Estimating theCustomer Lifetime: A Related Problem: Estimating the Customer Lifetime We also want to estimate the length of a customer’s lifetime If we know how long someone will remain a customer, we know when he will churn This can help us obtain an estimate of customer profitability Customers with a high lifetime value (profitability) are the ones we want to prioritize for retention efforts


Suppose Customers Have a 10% Chance of Leaving in the Next Month: Suppose Customers Have a 10% Chance of Leaving in the Next Month Probability the customer leaves in next month = .1  10% of customers have a lifetime of 1 month (round up) Probability the customer leaves in 2 months = .9 .1= .09  9% have a lifetime of 2 months (round up) Probability the customer leaves in 3 months = .9 .9 .1= .081  8.1% have a lifetime of 3 months (round up) Probability that customer leaves in x months = (.9)x1 .1


Average Customer Lifetime: Average Customer Lifetime In statistics, we refer to this as a geometric distribution If the churn rate per month is x%, the average lifetime is 1/x% months In this example, the average lifetime is 1/.1= 10 months An ideal attrition model says all customers leaving in the next month get a score of 1 or 100% all other customers get a score of 0 or 0%


Attrition Modeling Summary: Attrition Modeling Summary It is very important to manage attrition, especially in maturing industries Slowing attrition may be the cheapest way to maintain a critical mass of customers There are different types of attrition voluntary, forced, expected Related to attrition is the customer lifetime useful in calculating lifetime customer value


CRM Sources: CRM Sources The vast majority of these CRM slides has been borrowed, adapted, or reworked from one of the two sources below 1. Michael Berry and Gordon Linoff, Customer Relationship Management Through Data Mining, SAS Institute, 2000 2. Michael Berry and Gordon Linoff, Mastering Data Mining, John Wiley & Sons, 2000 I have also consulted Data Mining Techniques (Wiley, 1997) by Berry and Linoff, in preparing these slides


Decision Trees and Churn Modeling: Decision Trees and Churn Modeling Learn about decision trees what are they? how are they built? advantages versus disadvantages Case study from the cellular telephone industry problem definition looking at the results other techniques for modeling churn


Data Mining with Decision Trees: Data Mining with Decision Trees A decision tree is a set of rules represented in a tree structure Easy to understand how predictions are made Easy to build and visualize Can be used for binary or multiple outcomes Roots, nodes, leaves, and splits


Data Mining with Decision Trees--continued: Data Mining with Decision Trees--continued We build the decision tree using the training and test sets The decision tree allows us to make predictions understand why certain predictions make sense understand which variables are most important spot unexpected patterns There is one path from the root to any leaf A given data record is associated with a single leaf


A Path Through the Decision Tree: A Path Through the Decision Tree At each node, a decision is made--which variable to split These variables are the most important Each leaf should be as pure as possible (e.g., nearly all churns) All records landing at the same leaf get the same prediction A small example follows


A Decision Tree for Widget Buyers: A Decision Tree for Widget Buyers 1 yes 6 no 10 yes 3 no 11 yes 9 no 2 yes 3 no 8 yes Res NY Res = NY Age > 35 Age 35, then not a widget buyer Rule 3. If residence NY and age <= 35, then a widget buyer Adapted from (Dhar & Stein, 1997) No No Yes # correctly classified Accuracy = total # =


Building a Decision Tree: Building a Decision Tree We start at the root node with all records in the training set Consider every split on every variable Choose the one that maximizes a measure of purity ideally, all churners on left and non-churners on right For each child of the root node, we again search for the best split i.e., we seek to maximize purity Eventually, the process stops no good split available or leaves are pure


Building a Decision Tree--continued: Building a Decision Tree--continued The above process is sometimes called recursive partitioning To avoid overfitting, we prune the tree using the test set to prune is to simplify After the tree has been built, each leaf node has a score A leaf score may be the likelihood that the more common class arises in training and test sets


The Leaves Contain the Scores: The Leaves Contain the Scores Overall, the most common class is No 11,112 records (including Sam’s) are associated with this leaf 96.5% are not fraudulent Sam gets a score of .965 Yes 3.5% No 96.5% Size 11,112 


Scoring Using a Decision Tree: Scoring Using a Decision Tree This example shows the relationship between records in a table and decision tree leaves Remember, each record will fall in exactly one leaf


A Real Life Example: Predicting Churn: A Real Life Example: Predicting Churn Churn data from the cellular industry Begin at the root node The training set is twice as large as the test set Both have approximately the same percentage of churners (13.5%) The tree will be built using training set data only Training Test 13.5% 13.8% 86.5% 86.2% 39,628 19,814


The First Split: The First Split 13.5% 13.8% 86.5% 86.2% 39,628 19,812 Training set on left, test set on right 3.5% 3.0% 96.5% 97.0% 11,112 5,678 14.9% 15.6% 85.1% 84.4% 23,361 11,529 28.7% 29.3% 71.3% 70.7% 5,155 2,607 < 0.7% Handset Churn Rate < 3.8% 3.8%


The First Split--continued: The First Split--continued The first split is made on the variable “Handset Churn Rate” Handsets drive churn in the cellular industry So, first split is no surprise The algorithm splits the root node into three groups low, medium, and high churn rates If we look at the child on the far right, we see that the churn rate has more than doubled Far-right child has a lift of 29.3/13.8 = 2.12


As the Tree Grows Bigger, Nodesand Leaves Are Added: As the Tree Grows Bigger, Nodes and Leaves Are Added 13.5% 13.8% 86.5% 86.2% 39,628 19,814 3.5% 3.0% 96.5% 97.0% 11,112 5,678 14.9% 15.6% 85.1% 84.4% 23,361 11,529 28.7% 29.3% 71.3% 70.7% 5,155 2,607 = 0.18 Total Amt Overdue 67.3% 66.0% 32.7% 34.0% 110 47 70.9% 52.0% 29.1% 48.0% 55 25 = 88455


As the Tree Grows--continued: As the Tree Grows--continued CALL0 = calls in the most recent month CALL3 = calls four months ago CALL0/ CALL3 is a derived variable The number of calls has been decreasing Total Amt Overdue = total amount of money overdue on account When a large amount of money is due, the voluntary churn rate goes down It is easy to turn the decision tree into rules


Three of the Best Rules: Three of the Best Rules If Handset Churn Rate 3.8% AND CALL0/CALL3 < 0.0056 AND Total Amt Overdue < 88455 THEN churn likelihood is 52.0% (on test set) If Handset Churn Rate 3.8% AND CALL0/CALL3 < 0.0056 AND Total Amt Overdue < 4855 THEN churn likelihood is 66.0% (on test set) If Handset Churn Rate 3.8% AND CALL0/CALL3 < 0.18 THEN churn likelihood is 40.4% (on test set)


From Trees to Boxes: From Trees to Boxes Another useful visualization tool is a box chart The lines in the box chart represent specific rules The size of the box corresponds to the amount of data at a leaf The darkness of a box can indicate the number of churners at a leaf (not used here) We use HCR and C03 as abbreviations for Handset Churn Rate and CALL0/CALL3 Fill in three boxes in top right on page 228


From Trees to Boxes--continued: From Trees to Boxes--continued HCR 3.8% C03 < 0.18 C03 0.0056 HCR 3.8% C03 0.18% HCR < 0.7% 0.7% HCR < 3.8%


Finding a Good Split at a Decision Tree Node: Finding a Good Split at a Decision Tree Node There are many ways to find a good split But, they have two things in common splits are preferred where the children are similar in size splits are preferred where each child is as pure as possible Most algorithms seek to maximize the purity of each of the children This can be expressed mathematically, but it is not our focus here


Synonymous Splits: Synonymous Splits There are many ways to determine that the number of calls is decreasing low amount paid for calls small number of local calls small amount paid for international calls Different variables may be highly correlated or essentially synonymous Decision trees choose one of a set of synonymous variables to split the data These variables may become input variables to neural network models


Lurking Inside Big Complex Decision Trees Are Simpler, Better Ones: Lurking Inside Big Complex Decision Trees Are Simpler, Better Ones


Problem: Overfitting the Data: Problem: Overfitting the Data Amt Overdue < 88455 13.5% 13.8% 86.5% 86.2% 39,628 19,814 28.7% 29.3% 71.3% 70.7% 5,155 2,607 58.0% 56.6% 42.0% 43.4% 219 99 70.9% 52.0% 29.1% 48.0% 55 25 CALL0/CALL3 < 0.0056 Total  Good! The training set and test set are about the same  Good! The training set and test set are about the same  Good! The training set and test set are about the same X Ouch! The tree has memorized the training set, but not generalized


Slide234: Sometimes Overfitting is Evident by Looking at the Accuracy of the Tree Proportion Correctly Classified Number of Leaves


Pruning the Tree: Prune here This is a good model Error rate is minimized on unseen data Pruning the Tree Error Rate Number of Leaves Test data Training data


Connecting Decision Trees and Lift: Connecting Decision Trees and Lift A decision tree relating to fraud within the insurance industry is presented next Assume the analysis is based on test set data The cumulative gains chart consists of a series of line segments Each segment corresponds to a leaf of the tree The slope of each line corresponds to the lift at that leaf The length of each segment corresponds to the number of records at the leaf


Relationship Between a Cumulative Gains Chart and a Decision Tree: Relationship Between a Cumulative Gains Chart and a Decision Tree No (3692) 80.0% Yes (923) 20.0% No (1029) 69.5% Yes (452) 30.5% No (1405) 76.4% Yes (435) 23.6% No (1258) 97.2% Yes (36) 2.8% Base Policy All Perils (1481) Collision (1840) Liability (1294) Fault No (606) 58.2% Yes (436) 41.8% No (423) 96.4% Yes (16) 3.6% Policy Holder (1042) Third Party(439) Accident Area No (56) 45.9% Yes (66) 54.1% No (550) 59.8% Yes (370) 40.2% Rural (122) Urban (920) Adapted from Berry & Linoff (2000)


Decision Trees: A Summary: Decision Trees: A Summary It is easy to understand results Can be applied to categorical and ordered inputs Finds inputs (factors) that have the biggest impact on the output Works best for binary outputs Powerful software is available Care must be taken to avoid overfitting


Case Study: Churn Prediction: Case Study: Churn Prediction Predicting who is likely to churn at a cellular phone company The cellular telephone industry is rapidly maturing The cost of acquiring new customers is rather high A handset is discounted in exchange for a time commitment As the market matures, it becomes cheaper to retain than acquire customers Review the difference between voluntary, forced, and expected attrition


The Problem: The Problem This case study took place in a foreign country Outside of North America and Europe Mobil telephones are rather advanced in newly industrialized countries – used instead of fixed line technology The penetration of mobile phones in the U.S. was 25% in 1999


The Problem-- continued: The Problem-- continued The penetration of mobile phones in Finland was >50% in 1999 The company is the dominant cellular carrier in its market It has 5 million customers Four or five competitors each have about 2 million customers


The Problem-- continued: The Problem-- continued The company has a churn rate of about 1% per month It expects its churn rate to jump There are companies in this industry with churn rates of 4% per month The company wants to assign a propensity-to-churn score to each customer and put the resulting model into production ASAP They decided to develop the expertise in-house, using consultants The work took place over a two month period


Customer Base: Customer Base Segment # Customers % of Customers Churn Rate Elite 1,500,000 30 1.3% Non Elite 3,500,000 70 0.9% Total 5,000,000 100 1.1% Elite customers exceeded some threshold of spending in the previous year They have been around at least one year Why is their churn rate higher?


Data Inputs: Data Inputs Customer information file – telephone number, billing plan, zip code, additional services, Elite Club grade, service activation date, etc. Handset information – type, manufacturer, weight, etc. Dealer information – marketing region and size Billing history Historical churn rate information by demographics and handset type


The Model Set Uses Overlapping Time Windows: The Model Set Uses Overlapping Time Windows For the model set, August and September churners are used We want to use the model to make predictions for November


Solving the Right Business Problem: Solving the Right Business Problem Initial problem: assign a churn score to all customers Complicating issues new customers have little call history telephones? individuals? families? voluntary churn versus involuntary churn how will the results be used? Revised problem: by Sept. 20th, provide a list of the 10,000 Elite customers who are most likely to churn in October The revised problem invites action


Build a Number of Churn Models: Build a Number of Churn Models Numerous decision tree models were built It was discovered that some customers were taking advantage of the family plan to obtain better handsets at minimal cost These customers used the marketing campaigns to their own advantage Different parameters were used to build different models The density of churners in the model set was one important parameter (50% worked well)


Database Marketing and the Catalog Industry: Database Marketing and the Catalog Industry Catalogs are everywhere For many people, catalogs are the preferred method of shopping Their resemblance to magazines is intentional For many people, the reaction is the same Catalogs are very old the first Sears catalog was published in 1888 Catalogs flourished over the next half century railroad, coast-to-coast mail service, mass-produced goods


Catalogs are Part of the Retail Business: Catalogs are Part of the Retail Business Catalog retailers know their customers as well as their merchandise They communicate one-to-one and not just through branding and mass market advertising Ability to customize catalogs Closely related to B2C Web retailing Many companies (e.g., Eddie Bauer) have stores, catalogs, and a retail Web site


The Catalog Industry: The Catalog Industry $95 billion industry in U.S. (Catalog Age, 1998) Catalog sales are approximately 10% of all U.S. retail sales B2C Web retail sales are approximately 1% of U.S. retail sales (1999) Most catalogs now have an online presence It is becoming difficult to distinguish between Web retailers and catalog companies


What do Catalogers Care About?: What do Catalogers Care About? Merchandising how to display and promote the sale of goods and services? Layout where to place items in the catalog? which models to use? Response closer to the subject of this course personalization, customization, recommendation which of several pre-assembled catalogs should be sent to a specific customer based on an estimate of his response to each?


Example of a Big Catalog: Fingerhut: Example of a Big Catalog: Fingerhut Manny Fingerhut started database marketing in 1948 in Minneapolis sold seat covers on credit to new car buyers Fingerhut and specialty catalogs emerged over time As of the late 1990s $2 billion in annual sales 400 million catalogs/year (> 1 million/day) terabytes of transaction history customer and prospect file with 1,400 fields describing 30 million households


Different Catalogers Have Different Approaches: Different Catalogers Have Different Approaches Fingerhut aims for a less affluent audience “4 easy payments of $19.95” is a way to hide higher interest rates the most profitable customers don’t have to be the wealthiest Victoria’s Secret sees its catalog as part of its branding every mailbox should have one Eddie Bauer has a multi-channel strategy catalog, retail stores, and Web are equal partners Many are aimed at a very narrow audience


The Luxury of Pooled Data: The Luxury of Pooled Data Unlike other industries, the catalog industry promotes the sharing of data through a company called Abacus U.S. Air doesn’t know about my travels on TWA or Northwest Airlines But a catalog company (e.g., Eddie Bauer) can decide to send you a catalog because of a purchase you made from a competing catalog (e.g., L.L. Bean) As a consumer, you may have observed that a purchase from catalog A sometimes triggers the arrival of catalog B


Abacus and the World Wide Web: Abacus and the World Wide Web Abacus is the catalog industry infomediary 1,100 member companies maintains a database of over 2 billion transactions includes the vast majority of U.S. catalog purchases sells summarized data to member companies details of individual transactions are not revealed facilitates industry-wide modeling DoubleClick a leader in Web advertising, now owns Abacus indicates the convergence of online and off-line retailing


Another Source of Data: Another Source of Data Household data vendors companies that compile data on every U.S. household Acxiom, First Data, Equifax, Claritas, Experian, R.L. Polk hundreds of variables (demographic, lifestyle, etc.) database marketers use this data to assemble mailing lists Available data internal data on customers industry-wide data from Abacus demographic data from household data vendors


What Some Big Catalog Companies areDoing with Data Mining: What Some Big Catalog Companies are Doing with Data Mining Neural networks are being used to forecast call center staffing needs New catalog creation what should go into it? which products should be grouped together? who should receive the catalog? Campaign optimization


Eddie Bauer: Eddie Bauer Eddie Bauer is an interesting case They seek to maintain a unified view of the customer across three channels give the customer a unified view of Eddie Bauer They are integrating all channels for CRM 400 stores catalogs Web sales


Vermont Country Store (VCS): Vermont Country Store (VCS) Eddie Bauer and other large companies use CRM extensively What about smaller companies? VCS is a successful family-owned business VCS is a $60 million catalog retailer (late 1990s) VCS first used CRM to improve the targeting of their catalog to increase the response rate They used SAS Enterprise Miner to achieve a dramatic return on investment


Early Company History: Early Company History Founded by V. Orton in 1945, after he returned from WWII As a child, he had helped his father and grandfather run a Vermont country store His first catalog went out in 1945 or 1946 36 items 12 pages mailed to 1000 friends and acquaintances contained articles and editorials had the feel of a magazine


The Next Generation: The Next Generation Son Lyman Orton took over in the 1970s he remains as owner and chairman the CEO is not a family member Lyman focused on the catalog side of the business VCS grew by about 50% per year during the 1980s without data mining The catalog industry, as a whole, expanded rapidly during the 1980s


The Industry Matures: The Industry Matures As the catalog industry prospered, it attracted new entrants Catalogs began to fill every mailbox Response rates declined Growth rates declined The cost of paper and postage increased rapidly By 1995, more than one-third of catalog firms were operating in the red VCS was still profitable, but concerned


Vermont Country Store Today: Vermont Country Store Today They seek to “sell merchandise that doesn’t come back -- to people who do” $60 million in annual sales 350 employees Use Enterprise Miner to build response (predictive) models VCS sells the notion of a simple and wholesome (old-fashioned) New England lifestyle


Business Problem: Find the Right Customers: Business Problem: Find the Right Customers The VCS vision does not appeal to everyone Historical data on responders and non-responders is used to predict who will respond to the next catalog Focus on existing customers upside: we need to select a subset of a pre-determined set of customers downside: we are unable to identify outstanding new prospects The goal of response modeling is to find the right customers


Possible Objectives or Goals: Possible Objectives or Goals Increase response rate Increase revenue Decrease mailing costs Increase profit Increase reactivation rate for dormant customers Increase order values Decrease product returns


RFM: Common Sense in Retailing: RFM: Common Sense in Retailing Recency: customers who have recently made a purchase are likely to purchase again Frequency: customers who make frequent purchases are likely to purchase again Monetary: customers who have spent a lot of money in the past are likely to spend more money now Each one of these variables is positively correlated with response to an offer RFM combines all three variables


Where to Begin: Where to Begin Any of the previous goals can be addressed via data mining The first step is to define the goal precisely The specific goal determines the target variable Build the customer signatures orders placed by customer over time items purchased by category over time customer information from internal sources use customer zip code to add external data purchase household and neighborhood level data


RFM and Beyond: RFM and Beyond RFM uses recency/frequency/ monetary cells to determine who receives a catalog On left is the U.S. population (as of Nov. 1999) by age and gender In this section, we compare RFM and more sophisticated approaches


The Cells of the Table Below are Interesting: The Cells of the Table Below are Interesting More boys than girls until age 25 More females than males after age 25 75 – 79 year old men fought in WWII


Cell-Based Approaches are Very Popular: Cell-Based Approaches are Very Popular RFM is used often in the catalog industry Cell-based approaches are used to assess credit risk in the banking industry (e.g., CapitalOne) The insurance industry uses cell-based approaches to assess risk Market research also uses cell-based approaches, especially when the cells reflect demographics that can be used to predict purchasing behaviors


RFM is a Powerful General Methodology: RFM is a Powerful General Methodology Divide the “population” into cells based on known variables e.g., age, sex, income, bushiness of mustache Measure item of interest (e.g., response) for each cell (e.g., via a test mailing list) In full-sized mailing, focus on the cells of interest through selection, pricing, advertising


RFM: The Baseline Method for the Catalog Industry: RFM: The Baseline Method for the Catalog Industry RFM was developed in the 1980s, when the catalog industry was growing rapidly Proved to be very valuable, so it spread rapidly Became less effective in the 1990s at a time when costs (including paper costs) started rising rapidly Not a customer-centric approach, but important for comparison purposes


Sort List by Recency: Sort List by Recency Sort customers from most recent purchase date to least recent arrangement is based entirely on rank Break the list into equal-sized sections called quantiles top, bottom high, medium, low quartiles quintiles deciles This quantile becomes the R component of the RFM cell assignment


Sort List by Frequency: Sort List by Frequency Resort the list by frequency, break the list into quantiles But how do we measure frequency? total number of orders divided by the number of months since the first order average number of orders per month over the past year The quantile becomes the F component of the RFM cell assignment


Sort List by Monetary Value: Sort List by Monetary Value Resort the list by monetary value, break the list into quantiles But, what is the right measure? total lifetime spending average dollars per order The quantile becomes the M component of the RFM cell assignment Now, imagine a customer who has not placed an order in a little while doesn’t purchase that often when she does place an order, it tends to be big


From Sorted List to RFM Buckets: From Sorted List to RFM Buckets 341 Recency Frequency Monetary 1


RFM Cube: RFM Cube Recency = 5 Recency = 4 Recency = 3 Recency = 2 Recency = 1 $ = 5 $ = 4 $ = 3 $ = 2 $ = 1 Frequency = 5 Frequency = 4 Frequency = 3 Frequency = 2 Frequency = 1


RFM Cells are Not All the Same Size: RFM Cells are Not All the Same Size Consider the small example on left Do the numbers make sense?


Each Cell Has a Different Response Rate: Each Cell Has a Different Response Rate The better quantiles along each axis generally have higher response As expected, the 111 cell has the best response and cell 222 has the worst


Using RFM Cells to Determine Which Customers Should Receive a Catalog: Using RFM Cells to Determine Which Customers Should Receive a Catalog Given a customer list of one million Build a cube of 5 x 5 x 5 = 125 RFM cells Test mail to a random sample with all cells represented Given a budget that will cover test mail plus sending out 100,000 catalogs Mail catalogs to customers in top cells with respect to expected response to reach 100,000


RFM is a Testing Methodology: RFM is a Testing Methodology RFM involves testing the marketing environment in order to find the best cells direct test historical data Such testing makes good marketing sense However, it does not use a test set concept, the way we do in building a predictive model


Choosing Profitable RFM Cells: Choosing Profitable RFM Cells Calculate break-even response rate Example cost per piece = $1.00 revenue per response = $40.00 net revenue per response = $39.00 let r = response rate profit = 39r – (1 – r) profit > 0 39r > 1 – r 40r > 1 r > 2.5% mail to any cell with > 2.5% response in test mailing


Choosing Profitable Cells: Choosing Profitable Cells If we need a 2.5% response rate to break even, we choose cells 111 and 121 These account for 25% of the population


Using RFM Cells as Control Groups: Using RFM Cells as Control Groups VCS wanted to test a new catalog designed to appeal to 53-71 year olds The motivation was to exploit nostalgia for teenage and young adult years The new catalog included color images of products black and white images of 1950s and 1960s The test mailing compares target segment against others from the same RFM cell


Using Neural Networks for Response Modeling: Using Neural Networks for Response Modeling RFM is our baseline method RFM has a number of drawbacks At this time, we seek to gain a basic understanding of neural network models What are the pitfalls to be avoided? In this section, we discuss the neural network response model built by VCS


Problems with the RFM Approach: Problems with the RFM Approach RFM cell assignments are not scores is cell 155 better than cell 311? a test mailing is required RFM approach misses important segments Christmas shoppers: in November they may look like they have stopped buying from catalog there may be pockets of good responders in “bad” RFM cells there may be pockets of bad responders in “good” RFM cells


Problems with the RFM Approach -- continued: Problems with the RFM Approach -- continued Proliferation of cells there can be a large number of RFM cells RFM approach hits the same customers over and over the same people end up in the “best” RFM cells these customers suffer from “offer fatigue” RFM approach does not make use of some valuable data VCS knows where customers live (demographics) VCS knows about past behavior (purchases, returns, categories) VCS knows which customers are gift givers


Problems with RFM Approach -- continued: Problems with RFM Approach -- continued Focus on cells rather than individuals cell-based marketing treats everyone who falls into the same basket the same way because cell definitions tend to be simplistic, this doesn’t always make sense Predictive models are more powerful than profiles cell definitions are profiles profiling assigns a customer to a cell whose aggregate behavior is measured predictive models use input variables to predict future behavior of each customer


Back to VCS: Back to VCS After some initial success with RFM in the 1980s, VCS became less enchanted with the technique In 1998, they were ready to try something new VCS worked with SAS consultants to see if Enterprise Miner could achieve better results They compared RFM, regression, neural networks, and decision trees


Experimental Design: Experimental Design The model set data came from previous catalog mailings with data on responders and non-responders Using this data, VCS built RFM, regression, neural network, and decision tree models to predict who would respond to two other mailings that were also in the past, but more recent than the model set data October mailing November mailing Calculate expected response, revenues, and profit assuming models had been used


Available Data on Each Customer: Available Data on Each Customer Household ID Number of quarters with at least one order placed Flags indicating payment by personal check, credit card, or both Number of catalogs purchased Number of days since last order Number of purchases from the various catalog types Dollars spent per quarter going back several years Note that R, F, & M are each represented


How Models were Judged: How Models were Judged Compare regression, neural networks, and decision trees with RFM baseline with respect to percent increase in dollar sales Look at return on investment in data mining project How do we guess sales since the comparison is hypothetical? SAS and VCS were very clever


Using Historic Response Data to Compare the Models: Using Historic Response Data to Compare the Models Challenger models will pick a different mailing list than the one generated using the RFM approach How do we guess what someone would have spent had we sent her a catalog, given that we didn’t? Each model was used to score the list and choose its own top 1.4 million prospects for actual recipients, use actual dollars spent for non-recipients, use average spending of recipients with the same score


The Results: The Results The neural network model was the model selected by VCS The model predicted an increase in sales ranging from 2.86% to 12.83% perhaps, depending on the specific catalog The model yielded a return on investment of 1,182% More detailed information is confidential Let’s revisit the topic of neural networks


Neural Network Architecture: Neural Network Architecture output layer hidden layer input layer inputs are usually columns from a database


Inside Each Hidden Node: Inside Each Hidden Node Inputs are real numbers between -1 and 1 (data is transformed) Links have weights At each node, weighted inputs are summed and passed through a transfer function The output value is between 0 and 1 The output node functions similarly i1 i2 i3 w1 w2 w3 inputs output


Training a Neural Network: Training a Neural Network Given a pre-specified training set and random initial link weights observe one pattern (a single inputs/output combination) at a time iteratively adjust the link weights stop when final weights “best” represent the general relationship between inputs/output combinations This is called backpropagation An alternative approach is to use genetic algorithms


Procedure for Using Neural Networks: Procedure for Using Neural Networks Transform all the input variables to be between -1 and +1 and output variables between 0 and 1 Divide the data into training, testing, and validation sets Train the neural network using the training set until the error is minimized or the link weights no longer change Use the test set to choose the best set of weights for the model Compare models and predict performance using the validation set


Neural Networks: Pitfalls to be Avoided: Neural Networks: Pitfalls to be Avoided Decision trees can easily handle hundreds of input variables Neural networks cannot more input nodes  more link weights to be adjusted  the larger the training set  computationally burdensome one solution: build a decision tree first and use the variables that appear high in the decision tree in the neural network Categorical variables e.g., re-renter of Ryder trucks in the training set, define a variable that takes on the value 1 if the customer is a re-renter and 0 otherwise


Neural Networks: Pitfalls to be Avoided-- Continued: Neural Networks: Pitfalls to be Avoided-- Continued Be on the lookout for data skew caused by large outliers e.g., in looking at net worth, Bill Gates is assigned 1 and everyone else is assigned 0 differences between cardiologists, college professors, and the homeless would be obscured possible solutions: throw out outliers or perform transformations using logarithms or square roots


Final Thoughts on CRM: Final Thoughts on CRM Cross-sell models: see Chapter 10 in Berry and Linoff (Wiley, 2000) Data visualization: application to college selection CRM recap


CRM Sources: CRM Sources The vast majority of these CRM slides has been borrowed, adapted, or reworked from one of the two sources below: 1. Michael Berry and Gordon Linoff, Customer Relationship Management Through Data Mining, SAS Institute, 2000 2. Michael Berry and Gordon Linoff, Mastering Data Mining, John Wiley & Sons, 2000 I have also consulted Data Mining Techniques (Wiley, 1997) by Berry and Linoff, in preparing these slides