Mining Case Bases for Action Recommendation

1
Mining Case Bases for Action Recommendation

• Qiang Yang and Hong Cheng
• HKUST

2
Motivation

• Data Mining
• Given a dataset, can discover who is a good
  customer
• But
• What to do with a bad customer?
• What to do with a good customer whos turning
  bad?
• Our answer mine actionable plans

Data Mining
Actions
3
Example 1

• What to do to help Sammy get accepted to
  postgraduate school?

• Plan 1 (Dylan) improve Rec to above 4 AND GPA to
  3.6
• Plan 2 (Steve) improve TOEFL to 610

4
Example 2

• What to do to help Sammy to get loan approval?

• Plan 1 (Dylan) get higher income to 80K
• Plan 2 (Beatrice) get Married!!

5
Example 3

• How to prevent Sammy from leaving the mobile
  phone company?

• Plan 1 (Dylan) change Fee Reduction to 30, and
  increase free calls by 20 min
• Plan 2 (Beatrice) M? F !! ??

6
Action Recommendation Problem

• Recognize who are the (potential) negative-class
  members
• Segmentation Problem
• Recommend near optimal actions to help them
  switch to positive class
• Planning to achieve goals
• What does near-optimal mean?
• Cost
• Probability of success

Utilities
7
Our Solution

• First, discover who are the negative members
• May need to build classifiers using machine
  learning
• Second, discover goals or positive cases
• May need different role models for different
  negative members
• Need to balance
• Efficiency, Optimality, Cost
• Utilities
• Focus of this paper case mining

8
Case Based Reasoning Cycle

• Create
• Maintain
• Retrieve
• Revise
• Key Point how to find case bases?

9
Our Solution

• find cluster centroids
• find class boundary
• Need distance metric

10
Related Work

• Mining Optimal Actions for Intelligent CRM
  Ling, Chen, Yang, Chen, Industry Applications
  Paper (ICDM 02)
• Sequential Cost-Sensitive Decision Making with
  Reinforcement LearningPednault, Abe, and
  Zadrozny. (KDD'02)
• MetaCost A general method for making classifiers
  cost sensitive. P. Domingos. (KDD99)

11
Data Cleaning

• Taking a database with large number of
  attributes, many attributes are irrelevant
• We remove the irrelevant attributes using the
  OddsLogRatio Method
• Mlademnic and Grobelnik, 1999
• If Av occurs frequently in ve instances while
  infrequently in -ve instances, or vice versa,
  this attribute-value pair has discriminative power

12
Solution 1 Clustering ? Centroids
13
Solution 2 SVM ? Boundary Points
14
Case Retrieval and Adaptation

• x negative instance t positive instance
• p(t) the probability density around an
  instance t,
• cost(x, t) the cost of switch from x to target
  case t,
• maxCost the maximum value among the different
  costs of switching from x to every possible case
  y in the case base.

Cost of an attribute
15
Artificial Two-Class Data (Min Cost)

• Task
• comparing Centroid-based and SVM methods
• Conjecture
• performance depends on distribution ?
  well-separated versus closely mixed

16
Artificial Two-Class Data (Min Cost)
17
Artificial Two-Class Data (CPU Time)
18
UCI Data Set German Data
19
IBM Synthetic Data 100K records
20
Experiment on KDDCUP 98

• 479 attributes, 95,412 training instances

21
On-going Work Multi-step Planning

• Construct the association graph from the
  frequent subsequences.

22
Conclusions

• Objectives
• Case Base Mining from databases
• Purpose generating plans for CRM
• Algorithms
• SVM is OK for small, well separated data
• Centroids-based method is better for large,
  mixed-distribution data
• Utility based more realistic
• Future
• better utility models, apply to clustering
• Negative member identification