1 minute read

Overview

Early ML project on the California housing dataset — a classic setup where economic and geographic features (median income, population density, latitude/longitude, proximity to the ocean) are heavily correlated. OLS looks fine on training data but generalizes poorly.

The finding

VIF analysis flagged clear multicollinearity — several features exceeded VIF > 10, meaning they were largely redundant with linear combinations of others. Regularization was the right response.

Benchmarking ridge, lasso, and OLS with cross-validated RMSE:

  • Ridge won, delivering 12% lower RMSE vs. the OLS baseline
  • Ridge shrunk correlated coefficients toward each other rather than zeroing them out (lasso’s tendency), which suited the multicollinearity pattern
  • 4-member team; I led feature analysis and regularization strategy

Stack

  • Data: California housing (10+ demographic and economic features)
  • Preprocessing: missing-value imputation, categorical encoding, feature scaling
  • Diagnostics: VIF (variance inflation factor) for multicollinearity
  • Models: OLS, Ridge, Lasso via scikit-learn, tuned by cross-validation
  • Evaluation: cross-validated RMSE

What I’d change next

A nonlinear baseline (random forest or gradient boosting) would separate two questions the linear-only benchmark can’t distinguish: how much of the “regularization win” was really about handling multicollinearity vs. simply modeling nonlinearity better. Fixed effects on geographic clusters instead of raw lat/long would also help.

Completed during a summer ML internship at Prodigy Infotech, July–September 2023.