1 of 22

Association Rule Mining

  • Supervised Machine Learning Perspective
  • Generation of Association Rules
  • Course: Databases and Data Mining
  • Instructor: Jamolbek Mattiev

2 of 22

Learning Objectives

  • • Understand association rule generation
  • • Connect rule mining with supervised learning
  • • Learn evaluation metrics
  • • Interpret rule quality measures

3 of 22

Association Rule Mining Overview

  • • Discover relationships between variables
  • • Extract IF–THEN patterns
  • • Typically from transactional data

4 of 22

Supervised vs Unsupervised Context

  • • Traditional ARM is unsupervised
  • • Can be adapted for classification
  • • Class Association Rules (CARs)

5 of 22

Structure of Association Rule

  • A → B
  • Antecedent (IF part)
  • Consequent (THEN part)
  • Strength measured by metrics

6 of 22

Key Metrics

  • • Support
  • • Confidence
  • • Lift
  • • Conviction

7 of 22

Support Definition

  • Support(A→B) = P(A ∪ B)
  • Measures frequency of rule in dataset

8 of 22

Confidence Definition

  • Confidence(A→B) = P(B|A)
  • Measures reliability of rule

9 of 22

Lift Definition

  • Lift = Confidence / Support(B)
  • Lift > 1 indicates positive association

10 of 22

Rule Generation Process

  • 1. Find frequent itemsets
  • 2. Generate candidate rules
  • 3. Evaluate using confidence
  • 4. Prune weak rules

11 of 22

Class Association Rules (CAR)

  • Consequent restricted to class label
  • Used for classification tasks
  • Example: Symptoms → Disease

12 of 22

Relation to Decision Rules

  • • Similar IF–THEN structure
  • • Derived differently
  • • Based on frequency patterns

13 of 22

Threshold Selection

  • • Minimum support
  • • Minimum confidence
  • • Trade-off between quality and quantity

14 of 22

Overfitting in Rule Mining

  • • Too low thresholds → many weak rules
  • • Need validation
  • • Use statistical testing if needed

15 of 22

Evaluation on Test Data

  • • Apply generated rules to unseen data
  • • Measure classification accuracy
  • • Compare with other models

16 of 22

Advantages

  • • Interpretable rules
  • • Useful for knowledge discovery
  • • Flexible application domains

17 of 22

Limitations

  • • Large rule sets
  • • Redundancy
  • • Computational complexity

18 of 22

Applications

  • • Market basket analysis
  • • Medical diagnosis
  • • Fraud detection
  • • Recommendation systems

19 of 22

Discussion Questions

  • • How choose optimal thresholds?
  • • When use CAR instead of decision trees?
  • • How evaluate rule quality?

20 of 22

Summary

  • • Association rules reveal hidden patterns
  • • Rule generation depends on thresholds
  • • Can be adapted for supervised tasks
  • • Evaluation is essential for reliability

21 of 22

Support vs Confidence (Example)

22 of 22

Minimum Support vs Number of Rules