Access Transaction Data with Temporal Behavior Maps Shafi Rahman Director, Analytic Science FICO © 2014 Fair Isaac Corporation. Confidential. This presentation is provided for the recipient only and cannot be reproduced or shared without Fair Isaac Corporation’s express consent. Patent Pending Forward-Looking Statements Product roadmaps and similar marketing materials should be considered forward looking and subject to future change at FICO’s discretion. Future functionality, features or enhancements as shown are FICO’s current projections of the product direction, but are not specific commitments or obligations. 2 © 2014 Fair Isaac Corporation. Confidential. Transaction Data Is Every Where Improves Decisions Across Businesses Cash Grocery Utility Bill Cash Transfer Call Center Web Access Payment Credit Card Bill Time 3 © 2014 Fair Isaac Corporation. Confidential. Using Transaction Data For Business Decisions Predicting Attrition Using Credit Card Transaction Data βΊ Predicting card closure Masterfile Data (Model A) +Tx Data (Model B) 0.73 0.75 Recall 20.3% 24.3% Precision 8.4% 9.9% AUROC π πππππ = # ππππ ππ πππ ππ ππ π‘ππ 5% πππ‘ππ # ππ ππππ ππ πππ π ππππππ πππ = 4 # ππππ ππ πππ ππ ππ π‘ππ 5% # πππππ ππ ππ π‘ππ 5% © 2014 Fair Isaac Corporation. Confidential. Predicting Attrition Using Credit Card Transaction Data 5 Build Predictive Model with Transaction Data Deploy Model Begin Usage for Decisions (Total time to value) Traditional Transaction Approach 5 months 7 months 12 months Temporal Behavior Maps 3 months 2 months 5 months © 2014 Fair Isaac Corporation. Confidential. How To Improve Time To Value By Using A Reusable Transaction Variable Library Efficient Scalable Portable 6 © 2014 Fair Isaac Corporation. Confidential. βΊ Quickly compute transaction variables βΊ Consume new data feeds and data elements βΊ Calculate new characteristics βΊ Use same library in modeling and production βΊ Easy implementation Risk Prediction Using Masterfile Data The Behavior Risk Scores Are Quite Predictive Behavior Score 719 683 667 622 Jan 1 Feb 1 Mar 1 Apr 1 $500 Cash adv Overlimit Min payment made Min payment made Customer Behavior 7 © 2014 Fair Isaac Corporation. Confidential. Delinquent Bureau hit Timing of Transactions Contains Vital Information Which Masterfile Lacks Summarized Data Cardholder #1 Cardholder #2 8 © 2014 Fair Isaac Corporation. Confidential. Transaction Data Cash Merchandise Better and More Timely Risk Detection Using Transaction Data Behavior Score 719 683 667 622 Jan 1 Feb 1 Mar 1 Apr 1 722 Transaction Score 708 690 665 $500 Cash adv 620 Overlimit Min payment made Min payment made Customer Behavior 9 © 2014 Fair Isaac Corporation. Confidential. 600 580 Delinquent Bureau hit Challenges Working With Tx Data βΊ Iterative Computation βΊ Recursive Approximation 10 © 2014 Fair Isaac Corporation. Confidential. Computing Characteristics Historical Transaction Data of John Doe 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 # of fields vary from few to few dozens 11 © 2014 Fair Isaac Corporation. Confidential. πΆβππ1 = $ π ππππ‘ ππ πππ π‘ 30 πππ¦π $ π ππππ‘ ππ πππ π‘ 60 πππ¦π πΆβππ2 = $ π ππππ‘ ππ πππ π‘ 15 πππ¦π # of chars vary from 10 to 10,000 Computing Characteristics Using Iterative Approach πΆβππ1 = $ π ππππ‘ ππ πππ π‘ 30 πππ¦π $ π ππππ‘ ππ πππ π‘ 60 πππ¦π Compute on 03/01/2014 πΆβππ1 = 12 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 © 2014 Fair Isaac Corporation. Confidential. = $ π ππππ‘ ππππ 01/31 π‘π 03/01 $ π ππππ‘ ππππ 01/01 π‘π 03/01 $ 250 + $ 100 + $ 250 $ 50 + $20 + $520 + $ 120 + $630 + $ 250 + $ 100 + $ 250 = 0.309 Unable to leverage computation already done for numerator Accounting for New Transactions πΆβππ1 = $ π ππππ‘ ππ πππ π‘ 30 πππ¦π $ π ππππ‘ ππ πππ π‘ 60 πππ¦π 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 Compute on 03/07/2014 πΆβππ1 = = $ π ππππ‘ ππππ 02/06 π‘π 03/07 $ π ππππ‘ ππππ 01/07 π‘π 03/07 $ 100 + $ 250 + $ 780 + $ 40 $520 + $ 120 + $630 + $ 250 + $ 100 + $ 250 + $ 780 + $ 40 = 0.435 T = O(p), p = # of transactions Unable to leverage computation already done on 03/01/2014 13 03/07/14 Hotel Expense $ 780.00 03/07/14 Taxi $ 40.00 © 2014 Fair Isaac Corporation. Confidential. Handling More than One Characteristics πΆβππ2 = $ π ππππ‘ ππ πππ π‘ 15 πππ¦π 14 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 © 2014 Fair Isaac Corporation. Confidential. Compute on 03/01/2014 πΆβππ 2 = $ 100 + $ 250 = $ 350 T = O(n), n = # of variables Unable to leverage computation already done for Char1 Unable to Leverage Any Previous Computation πΆβππ2 = $ π ππππ‘ ππ πππ π‘ 15 πππ¦π 15 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 03/07/14 Hotel Expense $ 780.00 03/07/14 Taxi $ 40.00 © 2014 Fair Isaac Corporation. Confidential. Compute on 03/07/2014 πΆβππ 2 = $ 100 + $ 250 + $ 780 + $ 40 = $ 1170 Alternative Approach for Computing Characteristics Recursive Approximation Technique πΆβππ2 = $ π ππππ‘ ππ πππ π‘ 15 πππ¦π 16 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 03/07/14 Hotel Expense $ 780.00 03/07/14 Taxi $ 40.00 © 2014 Fair Isaac Corporation. Confidential. Compute on 03/07/2014 1 Char2 = ∗ Char2 03/01 +($ 40 + $ 780) λ Where λ is a decay factor How Recursive Approximation Technique Works? βΊ Consider: 1 ππ‘+1 = π₯πππ€ + ππ‘ λ ∞ ππ‘ = π=0 1 1 1 1 π₯ = π₯ + π₯ + π₯ + π₯ + β― 0 λπ π λ 1 λ2 2 λ3 3 1 1 1 1 ππ‘+1 = π₯πππ€ + π₯0 + 2 π₯1 + 3 π₯2 + 4 π₯3 + β― λ λ λ λ ππ‘+1 = π₯πππ€ + 17 © 2014 Fair Isaac Corporation. Confidential. 1 1 1 1 π₯0 + π₯1 + 2 π₯2 + 3 π₯3 + β― λ λ λ λ Limitations of Recursive Approximation Technique Not suitable for Suitable for Black Box Models Rules based solutions Triggers Interpretable values We need a third approach to overcome these limitations 18 © 2014 Fair Isaac Corporation. Confidential. Overcoming Challenges Using Temporal Behavior Maps (TBM) 19 © 2014 Fair Isaac Corporation. Confidential. Temporal Behavior Maps (TBM) A Simple Solution to Address the Challenges of Working with Tx Data Efficient Accurate Scalable 20 © 2014 Fair Isaac Corporation. Confidential. βΊ Computationally βΊ Palatable cheap and low latency values can be used across use cases βΊ Consumes new data feeds and data elements βΊ Easy to add new characteristic computation Portable βΊ Use Multiple modes βΊ Supports same library in modeling and production mode batch, incremental and real-time Temporal Behavior Maps Key Idea: Reuse Previous Computations βΊ Step 01/03/14 Restaurant $ 50.00 3 01/03/14 Movies $ 20.00 3 01/10/14 Jewelry $ 520.00 10 01/17/14 Clothes $ 120.00 17 01/24/14 Flight Booking $ 630.00 24 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 02/27/14 Car Rental $ βΊ Step 3 $ 70 10 $ 520 17 $ 120 34 24 $ 630 100.00 58 34 $ 250 250.00 58 58 $ 350 Aggregation step 1: Discretize time Period 21 © 2014 Fair Isaac Corporation. Confidential. 2: Store pre-computed summary of the transactions Base Map Computing Characteristics Using TBM Generation Step πΆβππ1 = $ π ππππ‘ ππ πππ π‘ 30 πππ¦π $ π ππππ‘ ππ πππ π‘ 60 πππ¦π 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 βΊ Use Base Map [3:$70, 10:$520, 17: $120, 24:$630, 34: $250, 58: $350] On 03/01/2014 πΆβππ1 = = $ π ππππ‘ ππππ ππππππ 31 π‘π 60 $ π ππππ‘ ππππ ππππππ 1 π‘π 60 $ 250 + $ 350 $ 70 + $520 + $ 120 + $630 + $ 250 + $ 350 = 0.309 22 © 2014 Fair Isaac Corporation. Confidential. Computing Multiple Characteristics Using TBM Much Cheaper Due to Base Maps πΆβππ2 = $ π ππππ‘ ππ πππ π‘ 15 πππ¦π 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 No need to traverse through Tx data again 23 © 2014 Fair Isaac Corporation. Confidential. βΊ Use Base Map [3:$70, 10:$520, 17: $120, 24:$630, 34: $250, 58: $350] On 03/01/2014 πΆβππ2 = $ π ππππ‘ ππππ ππππππ 46 π‘π 60 = $ 350 Accounting for New Transactions βΊ Only need to use the new transactions 01/03/14 Restaurant $ 50.00 3 01/03/14 Movies $ 20.00 3 01/10/14 Jewelry $ 520.00 10 01/17/14 Clothes $ 120.00 17 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 02/27/14 Cash Withdrawal 02/27/14 Car Rental βΊ Retrieve βΊ Update Old Base Map the Base Map 3 $ 70 3 $ 70 10 $ 520 10 $ 520 17 $ 120 17 $ 120 24 24 $ 630 24 $ 630 250.00 34 34 $ 250 34 $ 250 $ 100.00 58 58 $ 350 58 $ 350 $ 250.00 58 66 $ 820 Aggregation step 24 03/07/14 Hotel Expense $ 780.00 66 03/07/14 Taxi $ 40.00 66 © 2014 Fair Isaac Corporation. Confidential. Updating Characteristics Accounting for New Transactions πΆβππ1 = 25 $ π ππππ‘ ππ πππ π‘ 30 πππ¦π $ π ππππ‘ ππ πππ π‘ 60 πππ¦π 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 03/07/14 Hotel Expense $ 780.00 03/07/14 Taxi $ 40.00 © 2014 Fair Isaac Corporation. Confidential. βΊ Use updated Base Map [3:$70, 10:$520, 17: $120, 24:$630, 34: $250, 58: $350, 66:$820] On 03/07/2014 πΆβππ1 = = $ π ππππ‘ ππππ ππππππ 37 π‘π 66 $ π ππππ‘ ππππ ππππππ 7 π‘π 66 $ 350 + $ 820 $520 + $ 120 + $630 + $ 250 + $ 350 + $ 820 = 0.435 Defining Temporal Behavior Maps Domain specific A collection of base-maps 01/03/14 Restaurant $ 50.00 01/03/14 Movies $ 20.00 01/10/14 Jewelry $ 520.00 01/17/14 Clothes $ 120.00 01/24/14 Flight Booking $ 630.00 02/03/14 Car Repair $ 250.00 02/27/14 Cash Withdrawal $ 100.00 02/27/14 Car Rental $ 250.00 Price Base Map [3:$70, 10:$520, 17: $120, 24:$630, 34: $250, 58: $250] Cash Base Map [58: $100] Count Base Map [3:2, 10:1, 17: 1, 24:1, 34: 1, 58: 2] “Aggregator” function creates and updates the Base Maps 26 © 2014 Fair Isaac Corporation. Confidential. Key Concepts Base maps Temporal Behavior Map Aggregator Generator Period 27 © 2014 Fair Isaac Corporation. Confidential. βΊ Data structures, either maps or list of maps βΊ Store pre-computed summary of the transactions by period βΊ A collection of base maps for a given customer βΊ Summarizing βΊ Templates βΊ Time transaction data to generate base maps to generate characteristics from base maps discretization (daily, weekly, monthly,…) Modular Structure A Schematic Representation Derived Chars βΊRatio of spend in 1 and 2 months βΊSpend in last 1 month βΊSpend in last 2 months Total Spend Avg Spend Price Base Map 28 © 2014 Fair Isaac Corporation. Confidential. βΊRatio of cash in 1 and 2 months βΊCash in last 1 month βΊCash in last 2 months Total Cash Avg Cash Cash Base Map … … … … βΊRatio of visits in 1 and 2 months βΊVisits in last 1 month βΊVisits in last 2 months Total Visits Avg Visits Count Base Map Extensibility and Scalability Adding New Capabilities Is Easy Due to Modular Structure New data element New base maps New class of variables 29 © 2014 Fair Isaac Corporation. Confidential. Expand input data interface Extend aggregator to add a new base map Extend Generator to add a new function Working in Batch Mode Aggregator Sorted Tx Data Sorted on ID, date 30 © 2014 Fair Isaac Corporation. Confidential. Decision Engine Generators Temporal Behavior Maps Chars Decision Working in Real-Time or Incremental Modes Aggregator Temporal Behavior Maps Fetch NewTx Data Temporal Behavior Maps Storage TBM storage support random access 31 © 2014 Fair Isaac Corporation. Confidential. Decision Engine Generators Update Chars Decision Housekeeping Functions For Base Maps For Real-Time or Incremental Modes Maturation Trimming Append βΊ Use historical Tx data to instantiate base maps βΊ Discard earliest irrelevant periods as base maps grow with time βΊ Introduce a new base map in existing TBM store Point of singularity βΊ A date from which period is calculated 32 © 2014 Fair Isaac Corporation. Confidential. Portability The Biggest Contributor to the Accelerated Time to Value FICO® Blaze Advisor® FICO® Model Builder FICO® Analytic Offer Manager FICO® Analytic Cloud (coming soon) 33 © 2014 Fair Isaac Corporation. Confidential. TBM in Action: Use Case #1 Time to Event (TTE) Scorecards βΊTime to Event Model is a scorecard model that predicts propensity of an event to happen after a predetermined duration in future βΊ Combination of scorecards and survival models βΊ Predicts likelihood of each event to occur in a certain duration of time in future βΊ Ex. the probability that Guy Ritchie will use the card at a high end restaurant in the next 7 days βΊTTE considers more customer attributes at the individual customer level and includes a timing element βΊNeed to train one model for each event to be predicted 1000s of events to predict = 1000s of models to build 34 © 2014 Fair Isaac Corporation. Confidential. Training 1000s of TTE Model Highly Automated, Scalable, Distributed Modeling Platform Event-specific scorecards TBM makes it possible to compute 10,000 chars Transaction, Demographic and Product Data TBM Predictive modeling processes Event-specific training data Model for event B Auto-binning A Cleansing Summarization Sampling Feature Generation Data Reports 35 Model for event A © 2014 Fair Isaac Corporation. Confidential. Auto-variable selection Auto-training B Model for event C ... Auto-scoring C ... Model Reports Customer-Event propensities Using TTE Models For Making Timely and Relevant Offers TBM Makes It Easy to Deploy Variable Computation From Modeling to Production FICO® Analytic Offer Manager Objectives Decisions TTE Models TBM Data Propensity Matrix 36 Customer Offers Optimization Constraints Pr{Car Loan} Pr{Insurance} …. …. Pr{Restaurant} ... Jack 0.0815 0.0158 0.0439 0.2067 0.6564 0.0276 Joe 0.0906 0.0971 0.0382 0.0329 0.0363 0.0068 Jill 0.0127 0.0957 0.0766 0.2773 0.8499 0.0655 Pete 0.0913 0.0485 0.0795 0.0467 0.0934 0.0163 … 0.0547 0.0916 0.0646 0.1708 0.3923 0.0034 © 2014 Fair Isaac Corporation. Confidential. Propensity Matrix TBM in Action: Use Case #2 Multiple Decisions Supported in Production Using Tx Data FICO® Blaze Advisor® business rules management system Scoring Engine Transaction Scoring Core TBM Update TBM Storage* 37 © 2014 Fair Isaac Corporation. Confidential. Chars Attrition risk Bankruptcy Fetch Sorted Tx (Backend) Credit risk Revenue Decision fed into Client’s system Recapping Benefits of Temporal Behavior Maps Efficient Accurate Scalable 38 © 2014 Fair Isaac Corporation. Confidential. βΊ Computationally βΊ Palatable cheap and low latency values can be used across use cases βΊ Consumes new data feeds and data elements βΊ Easy to add new characteristic computation Portable βΊ Use Multiple modes βΊ Supports same library in modeling and production mode batch, incremental and real-time 39 Thank You! Shafi Rahman ShafiRahman@fico.com +1 (858) 353-8280 +91 (804) 137-1768 © 2014 Fair Isaac Corporation. Confidential. This presentation is provided for the recipient only and cannot be reproduced or shared without Fair Isaac Corporation’s express consent. Learn More at FICO World Related Sessions βΊBig Data Analytics Case Study: CIBC; Thursday, 1:30–2:30pm βΊMaking The Best Offer; Wednesday, 5:00–5:45pm Products in Solution Center FICO® Blaze Advisor® business rules management system FICO® Model Builder FICO® Customer Dialogue Manager FICO® Analytic Cloud Experts at FICO World βΊDr. Andrew Jennings βΊDr. Scott Zoldi 40 © 2014 Fair Isaac Corporation. Confidential. Please rate this session online! Shafi Rahman ShafiRahman@fico.com 41 © 2014 Fair Isaac Corporation. Confidential.