← Back to Research & Projects

Machine Learning for Data Science

Overview

Eight projects and a full set of course notes from Machine Learning for Data Science, a graduate course at USC. The projects run from k-nearest neighbours and regression through trees, SVMs, and clustering to semi-supervised learning and transfer learning. Each is a standalone notebook with its own data, built around preparation, model selection, and honest evaluation.

Projects

Project Methods Key result Open
KNN and Distance Metrics
Diagnose spinal abnormality from six biomechanical measurements.
K-nearest neighbours · Euclidean, Manhattan, Minkowski, Chebyshev, and Mahalanobis metrics · distance-weighted voting Test error 0.06 at k = 4; metric choice alone spans 0.06 to 0.17 Open
Linear Regression on Power Plant Output
Predict a power plant's hourly electrical output from ambient conditions.
OLS · polynomial and interaction terms · backward elimination · KNN regression Test MSE cut from 21.24 to 14.29 over the OLS baseline Open
Time Series Activity Classification
Recognize which of seven activities a person is doing from wireless sensor streams.
Time-domain features · logistic regression with RFECV · L1 penalties · naive Bayes 0.842 seven-class accuracy; the bending detector is exact on test Open
Regularization and Decision Trees
Make a clinical tree readable, then tame a 122-predictor crime regression.
Cost-complexity pruning · ridge · LASSO · principal component regression · XGBoost Ridge lowest of six models at 0.01732 test MSE Open
Imbalanced Classification with SMOTE
Detect failing air pressure systems in Scania trucks; failures are 1.67 percent of the training set.
Random forest · class weights · SMOTE · XGBoost with L1 Missed failures cut from 121 to 35 of 375 Open
Multi-Label SVM and K-Means Clustering
Label frog calls with family, genus, and species at once from 22 audio features.
Gaussian-kernel SVM one-vs-all · L1 linear SVM · SMOTE · k-means with silhouette selection All three labels right on 98.75 percent of test calls Open
Semi-Supervised and Active Learning
Measure how far models get when labels are scarce or chosen strategically.
Supervised SVM · self-training · k-means and spectral clustering · margin-based active learning Supervision holds the edge at 0.9708 over 30 fixed splits Open
Transfer Learning for Fungi Image Classification
Classify 9,114 microscopy images of fungal infections into five classes.
Frozen ImageNet backbones · ResNet50 and ResNet101 · EfficientNetB0 · VGG16 · DenseNet201 ResNet101 leads at 0.8476 test accuracy Open

Notes

I wrote these notes chapter by chapter through the course and assembled them into one book, from linear regression to deep learning. Every method gets its definitions, the formulas, a figure where one helps, and a plain explanation of why it works. Each chapter below opens the PDF at that chapter.

Ch Chapter Topics Open
1 Intro and Linear Regression Learning paradigms, the bias-variance trade-off, nearest neighbors, OLS and inference, multicollinearity, interaction effects Open
2 Classification Logistic and multinomial regression, Bayes' rule, LDA and QDA, naive Bayes, error measures, class imbalance Open
3 Linear Model Selection and Regularization Subset selection, cross-validation, ridge, lasso, elastic net, principal components and partial least squares regression Open
4 Tree-Based Methods Regression and classification trees, pruning, bagging, random forests, boosting and AdaBoost, Extra Trees, stacking and mixtures of experts Open
5 Support Vector Machines Maximal and soft margins, kernels, multi-class and multi-label SVMs, hinge loss, VC dimension, support vector regression Open
6 Unsupervised Learning K-means and k-medoids, hierarchical clustering, choosing K with silhouette and gap statistics, PCA and the SVD, Fisher's LDA Open
7 Semi-Supervised and Active Learning Self-training, co-training, semi-supervised SVMs, cluster and label, uncertainty sampling, query by committee Open
8 Neural Networks and Deep Learning The perceptron, MLPs and backpropagation, dropout and regularization, CNNs, transfer learning, RNNs, LSTM and GRU Open