Course for L2 Math students 6 sessions + 1 project.
Linear regression, logistic regression, MSE/MAE/R², precision/recall/F1/AUC, preprocessing (scaling, encoding, imputation), train/test split, k-fold cross-validation, sklearn pipelines.
Theory notebooks (tp_0): KNN (distance metrics, effect of k, validation curve, curse of dimensionality), Decision Trees (Gini/entropy, depth control, feature importance), SVM (maximum margin, soft margin C, kernel trick, RBF gamma), Bagging & Random Forests (bootstrap, variance reduction, OOB estimation, feature randomization), Gradient Boosting (AdaBoost, residual fitting, learning rate), Regularization (Ridge L2, Lasso L1 with feature selection, Elastic Net). TP1 — Regression: Model comparison on Diabetes dataset (baseline → KNN, Tree, SVR, RF, GB → GridSearchCV → cross-validation). TP2 — Classification: Model comparison on Wine dataset (scaling demo, baseline → 5 models → GridSearchCV on SVM → cross-validation).
Bias-variance tradeoff (empirical), Random Forest (bagging, n_estimators, feature importance), Gradient Boosting (learning rate, sequential correction), full model comparison with CV.
MLP with sklearn (architectures, activations), loss curves, GridSearchCV, RandomizedSearchCV, comparison with ensemble methods.
Tokenization, stopwords, Bag of Words, TF-IDF, text classification pipeline on 20 Newsgroups, model comparison (LogReg, NB, SVM), feature importance for text.
Pretrained Word2Vec exploration (analogies, similarity), document embeddings (mean pooling), embeddings vs TF-IDF comparison, conceptual intro to transformers.
Predict 24h temperature for 10 cities from 1000 cities x 2 weeks hourly data. Teams of 2-3, evaluated on ML-Arena. Grading: 50% performance + 50% presentation.
numpy
pandas
matplotlib
seaborn
scikit-learn