Computational data analysis

Summary

Two case studies applying rigorous, validation-first machine learning to two very different problems — predicting from more features than observations, and finding structure in noisy human biosignals.

The first case was a regression problem deep in the curse of dimensionality: 100 features against only 100 observations, where ordinary least squares overfits completely. We built an Elastic Net pipeline — KNN imputation for the roughly 15% of missing entries, robust scaling, one-hot encoding of categoricals — and tuned it inside a nested cross-validation so hyperparameters never saw the test fold. The model converged to a pure LASSO, keeping 26 of 115 coefficients and reaching an honestly estimated RMSE of 27.55.

The second case was a representation-learning study of the EmoPairCompete biosignals — heart rate, electrodermal activity, skin temperature — recorded from 26 people during a stressful team puzzle. The question was whether physiological stress is separable from rest, against the large differences between individuals. We compared PCA, CCA, t-SNE and UMAP across three standardization pipelines, validated throughout with leave-one-subject-out cross-validation so a person's identity could not leak across folds.

The findings were as much methodological as empirical. Inter-subject variability dominated — about 45% of variance between people versus 2% between task phases — so the stress signal only emerged after removing each person's baseline, lifting phase-prediction accuracy from 48% to roughly 57% and pointing to phasic electrodermal activity as the key channel. Non-linear embeddings added nothing over PCA, indicating an approximately linear signal, and in-sample canonical correlations overstated the physiology–emotion link by a factor of two until subject-level validation cut them down.

Across both cases the throughline was discipline about generalization: nested and leave-one-subject-out cross-validation, leakage-free preprocessing, and reporting the honest number rather than the flattering one.