pyGAM: balancing interpretability and predictive power using...
summary
pyGAM is a new pure‑Python implementation of pyGAM offering splines, smoothing penalties, per‑feature lambda tuning via generalized cross validation, scikit‑learn style fit/predict and grid/randomized search, plus model summaries and prediction intervals. The demo fits building power with weather, shows effective degrees of freedom, drops an uninformative cloud feature, and flags a p‑value bug. Missing: feature interactions and some diagnostics, but the library is built, usable, and reasonably comparable to mgcv and Prophet.
AI-written summary · about
chapters (21)
- 0:00Introduction and speaker background
- 1:33What is pyGAM and why create it
- 6:07Solving the PI data topless electricity problem
- 8:53What are generalized additive models
- 10:21Interpretability versus predictive power
- 11:44Understanding splines in GAMs
- 13:07Building nonlinear functions with splines
- 14:54Smoothing and the smoothing penalty
- 17:00Practical smoothing with pyGAM
- 18:29Cross-validation for smoothing parameters
- 19:30Selecting optimal lambda parameters
- 20:31Feature significance and selection
- 21:35Future features and community contributions
- 22:18Q&A: Comparing pyGAM with Prophet
- 24:02Q&A: Weather forecast usability for predictions
- 25:48Q&A: Multicollinearity in GAMs
- 26:51Q&A: Interpretability of GAM feature functions
- 28:16Q&A: Benchmarking against Gaussian processes
- 29:22Q&A: Handling non-Gaussian noise
- 30:25Q&A: Comparing pyGAM with mgcv
- 31:29Closing remarks and session summary
related talks
41mGeneralized Additive Models: Explainability Strikes Back
44mActually using GPs in practice with PyMC
01h 23mProbabilistic Graphical Models, HMMs using PGMPY
01h 30mMaking Gaussian Processes Useful
01h 33mMastering Gaussian Processes with PyMC
43mToward Human-Centered Machine Learning
18mFast Inference for Probabilistic Graphical Models
45mPyglmnet: A Python package for elastic net regularized generalized linear model