Winning with Simple, even Linear, Models
summary
Claims that linear models beat complex ones when you apply concrete tricks: add an interaction to solve XOR, use radial basis seasonal features and exponential sample weights to forecast a year ahead from ~1,100 points, tune with an evolutionary library (evil) and scikit‑learn pipelines, and deploy streaming via passive-aggressive streaming updates. Built systems include a SQL‑friendly co‑occurrence recommender, a realtime game‑skill estimator from match outcomes, and a hierarchical chicken weight model via rethinking/map2stan — all practical and working in production.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 4:43XOR with linear models
- 10:20Time series with RBFs
- 17:06Streaming models and PA
- 23:22Recommenders with probabilistic models
- 26:14Game skill with Bayesian updates
- 31:49Chicken weight models
- 35:40Conclusion
related talks
43mFeature Engineering for Time Series Forecasting
35mIntelligent Ruby: Getting Started with Machine Learning
45mFrom Ancient Greeks to Modern Geeks - Basic Machine Learning Algorithms
39mForecasting With Classical and Machine Learning Methods
45m"Breaking black-box machine learning"
41mGeneralized Additive Models: Explainability Strikes Back
22mReal-world applications of machine learning: Beyond the hype
40mMoving Forward Through The Darkness