DPPred: An Effective Prediction Framework with Concise Discriminative Patterns.

Area of Science:

• Machine Learning
• Data Mining
• Predictive Analytics

Background:

• Existing prediction models offer either strong interpretability (e.g., generalized linear models) or high effectiveness (e.g., tree-based models).
• A gap exists in models that provide both high accuracy and clear interpretability for complex datasets.

Purpose of the Study:

• To propose a novel Discriminative Pattern-based Prediction (DPPred) framework.
• To enhance prediction tasks by integrating the strengths of simple and complex modeling approaches.
• To achieve both high predictive accuracy and valuable interpretability.

Main Methods:

• DPPred utilizes concise discriminative patterns derived from prefix paths in tree-based models.
• It employs a search strategy to select optimal pattern combinations for generalized linear models.
• The framework integrates these selected patterns into a generalized linear model for prediction.

Main Results:

• DPPred demonstrates competitive accuracy compared to state-of-the-art methods across various scenarios.
• The framework provides valuable interpretability for developers and domain experts.
• In a clinical dataset case study, DPPred significantly outperformed baselines using a concise set of patterns.

Conclusions:

• DPPred offers an effective and interpretable solution for classification and regression problems.
• The framework successfully bridges the gap between predictive performance and model explainability.
• DPPred's pattern-based approach is particularly effective in scenarios with a vast potential feature space.