Guide towards algorithm explainability in machine learning
summary
Practical walkthrough that trains a loan-approval model on ~8,000 rows, gets 98% vs 54% when tested on production-like labels, and then applies concrete steps: imbalance fixes, permutation importance, anchors and counterfactuals via Alibi, and deployment of a model plus an explainer as a service on Seldon. Builds a UI and a “smile curve” for human-in-the-loop review and demonstrates the explainer pattern working end-to-end, with code at github.com/EthicalML/explainability-and-bias.
AI-written summary · about
chapters (8)
- 0:00Algorithm explainability intro
- 1:13Loan approval use case
- 6:17Why explainability matters
- 9:28Types of ML bias
- 12:57Data analysis for bias
- 18:18Model evaluation metrics
- 28:08Production monitoring with explainers
- 38:12Q&A
related talks
24mWhy and How to Build Explainability into your ML Workflow
44mImprove explainability of ML models to meet regulatory requirements
16mBuilding Trust through Explainable AI
35mWhy Do You Need an Explainable Ai?
01h 29mNavigating ML Explainability: From Theory to Implementation
41mMachine Learning Model Fairness and Interpretability with AWS
37mExplaining model explainability
47mDecoding the algorithm - why explainability and transparency matter when buildin...