Using structure to select features in high dimension
summary
Genomics-scale setting with ~1,000 samples and hundreds of thousands to millions of features, claim: using feature-structure improves selection and prediction. Implements a network-constrained lasso and a combinatorial min-cut/max-flow implementation that selects connected subnetworks; recovers 10 simulated causal features where unstructured tests and plain lasso fail. Reported RMSE progression: 0.21 (OLS) → 0.19 (lasso) → 0.14 (network lasso) → 0.10 (graph-selection then regression). Code and notebook on GitHub.
AI-written summary · about
chapters (6)
- 0:00Start
- 5:00Linear Regression
- 9:47Connectivity Constraint
- 10:31Network Constraint Lasso
- 14:52Train a Linear Regression
- 15:32Sparsity Constraint
related talks
25mAki720
27mFast, intuitive feature selection via regression on Shapley values
30mTraditional & Novel Feature Selection Approaches
38mWhy giving your algorithm ALL THE FEATURES does not always work
38mMixed and Covariate-Dependent Graphical Models
39mData cleaning: when less is more
01h 00mTensor Graphical Models For Complex And High Dimensional Data
33mNetwork Data Integration Enables Precision Medicine