Transcript Document
Two New Directions for Data Mining Charles Ling, PhD Department of Computer Science University of Western Ontario, Canada Director, Data Mining Lab, UWO [email protected] http://csd.uwo.ca/faculty/cling Two New Directions for Data Mining Action Mining Active Cost-sensitive Learning Charles Ling, PhD Action Mining for Profitable CRM Charles Ling, PhD Department of Computer Science University of Western Ontario, Canada Director, Data Mining Lab, UWO [email protected] http://csd.uwo.ca/faculty/cling CRM Customer Relationship Management: focus on customer satisfaction to improve profit Two kinds of CRM Enabling CRM: Infrastructure, multiple touch point, data integration and management, … – Oracle, IBM, PeopleSoft, Siebel Systems, … Intelligent CRM: data mining and data analysis – Vendors/products (http://www.kdnuggets.com/solutions/crm.html) Three Intelligent CRM Tasks Acquisition: direct marketing, application form, promotion methods, … Customization: cross/up-sale, segmentation, promotions, … Retention: Attrition/churn prevention Goal: through data mining to improve customer loyalty, satisfaction, and spending, resulting in increased company profits Action Mining Beyond model building and customer profiling Improve customer relationship: Actions changes What actions should you take to change customers from an undesired status to a desired one – – – – – From churn to loyal From inactive to active From low spending to high spending From non-customers to customers … and make the maximum profit (the ultimate goal) Research Issues Bounded Action Problem (BAP) – Types of actions are limited to k – How to find k action types to maximize profit The problems are NP-hard – Exponential to k Our solutions: heuristic/greedy search based on decision trees – Proactive Solution Charles Ling, PhD How Proactive Solution Works 1. Get Customer Data (marketing DB) 2. Build Customer Profiles 3. Search Actions for Maximal Profit 4. Action Delivery Step 1: Get Customer Data Marketing DB: Segmentation, data preparation, pre-processing… Define a “target”: undesired status and desired status ID Name Age Sex Service Rate Prof … Retained (Target) 1001 John 50 M H L A … Yes 3010 Sue 25 F M H D … No … … … … … … … … 40 M M H B … ??? … 1112 Jack Step 2: Build Customer Profile on target Automatically by Proactive Solution with probabilities on the target Service M H L Sex F Rate M L H Prob=0.1 Prob=0.9 Prob = 0.2 Prob=0.8 Prob=0.5 Step 3: Search Actions for Maximal Profit Proactive Solution searches more desired nodes in the profile… ID Name Age Sex Service Rate Prof … Retained … … … … … … … … … 40 M M H B … ??? 1112 Jack Jack: …, Service = M, Sex = M, Rate = H, … Profit =$4000 Service M H L Sex F Rate M L Prob=0.1 H Prob gain = -0.1 Serv: MH E.Profit= -400 Rate: HL Cost= $500 E.Net Profit= -900 Prob=0.9 Prob gain = 0.7 E Profit= $2800 Cost = E Net Profit= - Prob = 0.2 Prob=0.8 Prob gain = 0.6 E Profit=$2400 E.Profit=$2400 Cost=$800 E NetProfit=$1600 E.NetProfit=$1600 Prob=0.5 Prob gain = 0.3 E Profit=$1200 Cost=$400 E NetProfit=$800 Step 4: Action Deployment ID Name Prob Actions diff 1112 Jack … 0.6 3010 Sue 0.5 3421 Bill … Action costs Service: M H $800 Rate: H L SigAcc: 0 1 $500 Service: L M N/A NetProfit … $1600 … $700 $0 • Selective deployment: human intelligence, … • Customer segmentation by actions Reporting – on the web Advanced Features Accurate probability estimations Better evaluation methods – AUC of ROC Hard vs soft attributes – search many trees Beam-search Action correlation Charles Ling, PhD Case Study: Mutual Fund An insurance company selling mutual funds Task 1: For the current fund owners, how to improve their fund purchasing (from low to high spending)? Task 2: Some representatives are good performers but some are not; how to change bad performers to be good performers? Task 3: Many customers do not currently own mutual funds. How to market to them to buy mutual funds? Summary From model building to action mining (deployment) Business oriented: maximal net profit Proactive Solution: effective intelligent CRM Technically sophisticated Massive one-to-one customization Effective marketing and segmentation tool Charles Ling, PhD