Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong...

16
Post-processing Decision Post-processing Decision Trees Trees to Extract to Extract Actionable Knowledge Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department of Computer Science University of Western Ontario, Canada

Transcript of Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong...

Page 1: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Post-processing Decision TreesPost-processing Decision Treesto Extractto Extract

Actionable KnowledgeActionable Knowledge

Qiang Yang and Jie YinHKUST, Hong Kong

China

andCharles X. Ling and Tielin Chen

Department of Computer ScienceUniversity of Western Ontario, Canada

Page 2: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

CRMCRMCustomer Relationship Management:

focus on customer satisfaction to improve profit

Two kinds of CRM Enabling CRM: Infrastructure, multiple touch point

management, data integration and management, … Oracle, IBM, PeopleSoft, Siebel Systems, …

Intelligent CRM: data mining and analysis, database marketing, customization Vendors/products (see later)

Page 3: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

From Data Mining to ActionsFrom Data Mining to ActionsWhat to do to help Sammy to get loan

approval?

•Action 1 (Dylan): get higher income to 80K

• Action 2 (Beatrice): get Married!!

Income Married Cars Approved?Sammy 50K n 1 ?Beatrice 50K y 1 YesDylan 80K n 2 YesMathew 30K n 1 NoLarry 40K n 0 No

Basil 80K n 1 Yes

Customer Database

App

lican

ts

Page 4: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Actionable vs. Passive Data Actionable vs. Passive Data MiningMining

Improve customer relationship Actions (promotion, communication) changes

What actions to 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-buyers to buyers …

and still make a profit (the ultimate goal) Approach: Post-processing Decision Trees

Mining actions from decision trees Bounded action problem Bounded segment problem Our solutions

Page 5: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Post-processing Decision TreesPost-processing Decision Trees

1. Get Customer Data (marketing DB)

2. Build Customer Profiles

3. Search Actions for Maximal Profit

4. Action Delivery

Page 6: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Step 1: Get Customer DataStep 1: Get Customer Data

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

… … … … … … … … …

1112 Jack 40 M M L B … ???

Marketing DB: Segmentation, data preparation, pre-processing…Define a “target”: undesired status and desired status

Page 7: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Prob=0.1

Prob = 0.2 Prob=0.9 Prob=0.5

Service

RateSex

M L H

MF HL

Prob=0.8

Step 2: Build Customer Profile on targetStep 2: Build Customer Profile on targetAutomatically by Proactive Solution with probabilities on the target

Page 8: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Step 3: Search Actions for Maximal Step 3: Search Actions for Maximal ProfitProfit

ID Name Age Sex Service Rate Prof … Retained

… … … … … … … … …

1112 Jack 40 M M L B … ???

Proactive Solution searches more desired nodes in the profile…

Page 9: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Prob gain = 0.6E.Profit=$2400Cost=$800E.NetProfit=$1600

Prob gain = -0.1E.Profit= -400Cost= $500E.Net Profit= -900

Prob gain = 0.7E Profit= $2800Cost = E Net Profit= -

Prob gain = 0.3E Profit=$1200Cost=$400E NetProfit=$800

Prob gain = 0.6E Profit=$2400Cost=$800E NetProfit=$1600

Jack: …, Service = M, Sex = M, … Profit =$4000

Prob = 0.2 Prob=0.9 Prob=0.5

Prob=0.1

Service

RateSex

M L H

MF HL

Prob=0.8

Serv: MHRate: ? L

Page 10: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Step 4: Action DeploymentStep 4: Action Deployment

ID Name Prob diff

Actions Action costs

NetProfit

1112 Jack … 0.6 Service: M H

Rate: L M$800 … $1600

3010 Sue 0.5 SigAcc: 0 1

Service: L M$500 … $700

3421 Bill … N/A $0

• Selective deployment: human intelligence, … • Customer segmentation by actions

Page 11: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Practical Issue: Resource is Practical Issue: Resource is BoundedBounded

Limited number of account managers Thus, the number of customer segments is bounded Research: how to generate no more than K customer

segments, such that for each segment, find a set of common actions to apply

We call this the bounded segmentation problem (BSP)

Limited number of marketing actions Thus, types of actions are limited We call this the Bounded Attribute Set Problem (BASP)

Both problems are NP-hard.

Page 12: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

The Bounded Segmentation The Bounded Segmentation ProblemProblem

Resources are bounded! Group (potential) negative-class customers into pre-specified

k customer segments. Recommend “near optimal” actions to help each of the k

customer segments switch to a more profitable positive class. Each segment is applied by the same actions (same manager) The expected net profit is to be maximized

Each action may have a different cost and bring different profits The Bounded Segmentation Problem is NP-Complete

Equivalent to maximum coverage problem. NP-hard problem!

We seek approximate solutions!

Page 13: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

The Bounded Segmentation The Bounded Segmentation Problem: Greedy AlgorithmProblem: Greedy Algorithm

1. Discover who are negative-class customers. build decision tree as the classifier

2. Group negative-class leaf nodes into k customer segments using greedy algorithm.

Each customer segment one action set The total profit gain by applying such k action sets

can be maximized. Algorithm is based on finding the current largest

coverage in linear time

Page 14: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

An Example: K=2An Example: K=2

L 1 L 2 L 3 L 4

S ta tu s R a te

S e rv ice

L H

A B C D

0 .9 0 .2 0 .8 0 .5

If we want to find two customer segments (k=2) It is more profitable to transform L2L1 and L4L3 than others Profit gain = (0.9-0.2)*1-0.2 + (0.8-0.5)*1-0.1=0.7.

cost

Page 15: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

Experiment on Mutual Fund DataExperiment on Mutual Fund Data

GreedyBSP can find k customer segments with maximal profit. Result is very close to those found by OptimalBSP.

GreedyBSP is more scalable than OptimalBSP.

Page 16: Post-processing Decision Trees to Extract Actionable Knowledge Qiang Yang and Jie Yin HKUST, Hong Kong China and Charles X. Ling and Tielin Chen Department.

SummarySummary

From decision-tree model building to extracting actions for profit

Goal: maximal net profitResource is bounded

Design optimization solutions for action extractionBASP and BSP

Future: explore more efficient solutions