DS357 – Explainable AI (XAI) Project

Research Replication and Extension

This project is part of the DS357 course focused on Explainable AI (XAI).

Selected Paper

InstaSHAP: Interpretable Additive Models Explain Shapley Values Instantly
(ICLR 2025) :contentReference[oaicite:0]{index=0}

Problem Statement

SHAP is a powerful method for feature attribution in machine learning models.
However, computing exact SHAP values is very slow and computationally expensive.
This makes it difficult to use in real-time applications.

Objective

• Understand the selected research paper deeply
• Reproduce the original experiments
• Identify limitations in the approach
• Propose and implement an improved solution
• Compare results with the original method

Dataset Used

• Standard tabular benchmark datasets
• CUB (Caltech-UCSD Birds) dataset

XAI Method

• Original Method: Generalized Additive Models (GAM)
• Proposed Method: Explainable Boosting Machine (EBM)

Key Idea

The original paper uses GAM to approximate SHAP values quickly.
However, GAM assumes feature independence and ignores interactions.

Proposed Improvement

We replace GAM with EBM, which can model pairwise feature interactions.
This improves explanation quality while maintaining interpretability.

Implementation Details

• Model replication from the research paper
• Implementation of XAI surrogate model
• Training on the same datasets
• Evaluation using accuracy and speed metrics

Expected Outcomes

• Faster SHAP value approximation
• Better handling of feature interactions
• Improved explanation accuracy
• Real-time applicability

Project Phases

• Phase 1: Paper selection and proposal
• Phase 2: Experiment replication
• Phase 3: Research gap and extension

Repository Structure

• Phase_1_work: Paper analysis
• Phase_2_work: Replication code
• Phase_3_work: Improvement and results

Reproducibility

• Dataset links provided
• Requirements file included
• Instructions to run code
• Random seed for consistency

Authors

Team of 5 students – DS357 Course Project