Explaining Machine Learning Models with LLMs
summary
Claims that LLMs can explain both interpretable models and black boxes by taking model components, producing specific outputs, and finding real issues: uses GAM/EBM function-to-text conversion for modular interpretables and SHAP scatter plots for arbitrary models, sends per-feature plots plus dataset context (about 700 bike-rental samples) to an LLM, uses a chain-of-thought split with structured JSON outputs so the pipeline returns descriptions, highlighted anomalies, rules to check, notes on privacy/cost tradeoffs, and suggests how to validate results.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 3:00Why explainable AI is needed
- 6:01Challenges in explainable AI
- 7:08Talk to EBM and LLM summarization
- 10:00Shapley values and their role
- 12:53Generating scatter plots for features
- 14:02Sending plots to LLM for analysis
- 15:49Chain of reasoning for LLM accuracy
- 19:09Risks and limitations
- 22:05Optimizing cost and privacy
- 22:48Evaluating LLM explanations
- 24:23Improving metadata with LLMs
- 25:28Q&A
related talks
46mExplaining a Machine Learning blackbox
35mReading the Mind of an LLM
50mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
34mInterpreting Machine Learning Models: Why and How!
29mTowards Explainable Machine Learning
22mIntegrating LLMs with Traditional Data Analysis
01h 29mNavigating ML Explainability: From Theory to Implementation