DPClass: An Effective but Concise Discriminative Patterns-Based Classification Framework.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Traditional pattern-based classification methods often struggle with pruning non-discriminative patterns.
• Tree-based models excel at capturing feature interactions and handling diverse data types.

Purpose of the Study:

• To develop a novel classification framework integrating discriminative patterns from tree-based models with generalized linear models.
• To improve accuracy, interpretability, and prediction speed in pattern-based classification.

Main Methods:

• Extracting discriminative patterns as prefix paths from tree-based models (e.g., random forest).
• Compressing patterns using generalized linear models to select the most effective combinations.
• Implementing the Discriminative Pattern-based Classification (DPClass) framework.

Main Results:

• DPClass achieves performance comparable to state-of-the-art algorithms.
• High accuracy is maintained even with a limited number of top discriminative patterns (e.g., top-20).
• The framework offers significant interpretability and fast prediction capabilities.

Conclusions:

• DPClass provides an effective and interpretable approach to pattern-based classification.
• The method demonstrates the potential of combining tree-based and linear models for complex classification problems.
• DPClass offers a concise and highly explanatory framework for expert analysis.