EvoXplain framework reveals inconsistent ML model explanations

Researchers have developed EvoXplain, a new framework designed to assess the consistency of explanations generated by machine learning models. The tool investigates whether different training runs and model selection processes lead to similar underlying logic or divergent mechanisms, even when predictive performance is high. EvoXplain analyzes the structure of explanations themselves, rather than aggregated predictions, to reveal when a consensus explanation might not reflect any single trained model's reasoning. Initial evaluations on cancer genomics data using logistic regression and gradient-boosted trees showed that while accuracy remained high, the explanatory structures varied significantly across different training pipelines and even within simple tuning steps. AI

IMPACT Highlights the need for interpretability methods that account for mechanistic multiplicity in ML models, particularly in scientific applications.

RANK_REASON The cluster contains a research paper detailing a new framework for analyzing machine learning model explanations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Chama Bensmail Bensmail
• EvoXplain
• Gradient Boosted Trees
• logistic regression model
• The Cancer Genome Atlas

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →