Reproduce “FORCE: Feature-Oriented Representation with Clustering and Explanation” on OpenResearch

About this paper

Rishav Mukherjee, Jeffrey Ahearn Thompson

The FORCE framework integrates SHAP values into a deep neural network by using them to cluster observations and to initialize the attention layer, aiming to enhance predictive performance by leveraging feature importance information. This approach consistently improved accuracy, precision, recall, F1-score, and AUC across early-stage diabetes risk, heart disease, and credit approval datasets, with F1-score increases of up to 11 percentage points over baseline models.

Key insights

• Clustering observations based on their SHAP values can uncover latent data structures that are directly relevant to the outcome and enhance predictive power, outperforming clustering on raw features.
• Integrating explainable AI (XAI) techniques like SHAP values directly into a model's architecture, both for feature representation and attention initialization, can improve the model's learning and discriminative capabilities.
• Initializing attention mechanisms with SHAP values provides the neural network with a guided starting point for feature focus, potentially leading to more efficient learning and better generalization.

Summary by alphaXiv.