Synergizing model and human decision-making using augmented machine learning
summary
Models predict underwriter decisions rather than defaults, feature values and descriptions appear in the dashboard, and per-decision SHAP explanations show which features push approval up or down. Business rules come from depth‑1 scikit‑learn trees exported to a JSON config used as a rule engine, yielding a 60% reduction in manual review work. Experiments include monotone constraints in XGBoost, PSD2 transaction aggregation, and a challenge workflow that stores operator disagreements for retraining and feature improvements.
AI-written summary · about
chapters (22)
- 0:00Augmented ML collaboration
- 1:12Speaker background
- 2:15Florence's lending process
- 5:47Modeling Patrick's decisions
- 8:15Modeling human intuition challenges
- 9:41Building trust in ML models
- 10:23Features for transparency
- 11:47SHAP values for explanations
- 15:37Stakeholder feedback challenges
- 16:01Validating decision challenges
- 18:06Monotone constraints for explainability
- 19:30Model improvement feedback
- 20:13Semi-automating client review
- 22:02Data-driven business rules
- 24:09Decision trees for automation
- 26:14Automated review efficiency
- 27:17Future model directions
- 27:57Key takeaways
- 28:43Q&A: Economic changes
- 30:07Q&A: Business rule adjustments
- 31:11Q&A: Monotonic features
- 32:35Q&A: Rule monitoring
related talks
43mToward Human-Centered Machine Learning
35mSHAPtivating Insights: unravelling blackbox AI models
42mAugmenting Human Decision Making with Data Science
27mInterpretability and explainability in machine learning
31mBeyond the Black Box: Interpreting ML Models with SHAP
18mSynthesizing human and machine capabilities
38meXplainable Predictive Decisioning
13mHuman-in-the-Loop MLOps: Boost Model Performance by 40%