FORCE: Feature-Oriented Representation with Clustering and Explanation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning research actively explores latent structures to enhance predictive model accuracy.
• Current methods often cluster features to infer latent structures, but benefits can be limited with complex models.

Purpose of the Study:

• To propose FORCE (Feature Oriented Representation with Clustering and Explanation), a novel deep learning framework utilizing SHAP values.
• To improve predictive model performance by integrating latent feature representation and attention mechanisms guided by SHAP values.

Main Methods:

• FORCE employs a two-stage SHAP value integration: latent embedding for clustering and an attention mechanism.
• It guides neural network training by clustering absolute SHAP values and uses this latent information for attention.

Main Results:

• FORCE demonstrated significant performance improvements across three real-life datasets compared to baseline networks.
• For instance, the F1 score for Polycystic Ovarian Syndrome detection improved from 0.80 to 0.99.

Conclusions:

• The FORCE framework effectively enhances deep learning by leveraging SHAP values for latent pattern discovery and attention.
• This approach boosts overall discriminative capability and predictive accuracy in machine learning models.