Konstantin Vorontsov

A survey lecture: machine learning as formulations of optimisation problems, a model, a loss function, a regulariser. Regression and classification, continuous approximations of the threshold loss function, SVM and logistic regression, quantile, robust and SVM regression; the L0, L1, L2, Elastic Net, SFM, RFM regularisers and feature selection; learning to rank: ordinal regression, the pairwise approach and AUC, LambdaRank; estimating a density and a mixture of distributions, clustering as a special case of a mixture, the one-class SVM; semi-supervised learning, the transductive SVM, PU-learning; matrix factorisations, autoencoders, multidimensional scaling, graph factorisations, GraphEDM; transfer learning, multi-task and self-supervised learning, distillation, learning using privileged information, generative adversarial networks; a typology of machine learning problems.

The combinatorial theory of overfitting: the probability of overfitting and why the Vapnik–Chervonenkis bounds are loose; experiments with overfitting — a monotone chain of algorithms, the effect of splitting and connectivity of the family on the probability of overfitting and on the complete cross-validation estimate, overfitting of chains, the exact formula for the probability of overfitting when choosing between two algorithms; the splitting–connectivity graph of a family of algorithms, examples for linear classifiers, a monotone chain and a two-dimensional net; generating and forbidden sets, upper connectivity and inferiority of an algorithm; the splitting–connectivity upper bound, the idea of its proof and its properties.

Interpretability and explainability of machine learning models: goals, tasks and basic notions, who needs explanations and why; interpretable models; feature importance in linear regression: the t-statistic, net effects, contributions and situational importance; partial dependence plots (PDP) and individual conditional expectation plots (ICE); permutation feature importance; the Shapley value from cooperative game theory, its axioms and Monte Carlo estimation, its use in linear regression; surrogate modelling in the neighbourhood of an object, the LIME, Anchors, SHAP, Shapley Kernel and SAGE methods; the Shapley value for training objects, Data Shapley and Gradient Shapley; counterfactual explanations.

Active learning: the model itself chooses which objects to give the teacher for labelling. The problem setting, why active learning is faster than passive learning, learning curves; pool-based sampling, query synthesis and stream-based selective sampling, cost-sensitive methods; uncertainty sampling (least confidence, minimum margin, maximum entropy), query by committee, version space reduction, expected model change, expected error reduction; query synthesis by the Nelder–Mead method and by the variance reduction criterion; density weighting; exploratory actions, the ε-active and EG-active algorithms with the exponentiated gradient; active learning in crowdsourcing: reconciling annotators' labels, models of their reliability and of object difficulty, assigning tasks to annotators.

Reinforcement learning as learning from the environment's response: the multi-armed bandit problem, estimating the value of an action, the greedy and ε-greedy strategies, the exploration–exploitation trade-off, the UCB method, the exponential moving average, the pursuit method and reinforcement comparison; an environment with states, the Markov decision process, how it differs from ordinary machine learning problems, the discounted return, the state and action value functions, the Bellman equation, the SARSA and Q-learning methods, on-policy and off-policy; deep Q-learning DQN on the example of Atari games, experience replay; gradient optimisation of the policy (policy gradient) and its connection with log-likelihood maximisation; model-based and model-free approaches, a regression model of the reward with an upper confidence bound, LinUCB, recommending news articles, evaluating a policy on historical data.

Learning on a data stream: the formulation of the online learning problem, the learning curve, the issues: efficient model updates, making the model more complex as the sample grows, the equivalence of online and offline, storing the sample and catastrophic forgetting; how online learning differs from incremental learning, continual learning, anytime algorithms; lazy learning and online prototype selection; the online naive Bayes classifier with exponential distributions of features; the perceptron, Passive-Aggressive and online gradient descent; the recursive least squares method and the Sherman–Morrison formula; incremental training of decision trees ID5R; online training of an ensemble by the Hedge(β) algorithm and its financial interpretation; online deep learning and Hedge BackProp; incremental learning of new classes, iCaRL.

Time series forecasting problems: the time series model, the forecasting horizon, the least squares method; econometrics as the main source of problems — trends, seasonalities, change points; examples of forecasting sales volumes and electricity prices; the linear autoregression model; a brief survey of forecasting methods; adaptive short-term forecasting methods: the exponential moving average and the smoothing parameter, the models of Holt, Theil–Wage, Winters with linear and exponential trend, adaptive autoregression; the tracking signal and a statistical test of model adequacy, the Trigg–Leach model; adaptive selection and adaptive composition of models, the LAVR method — local adaptation of weights with regularisation, the monotonicity constraint as a regulariser; experiments on the sales forecasting problem.

Association rule mining as unsupervised learning: binary features, support and frequent itemsets, the association rule, minimum support MinSupp and minimum confidence MinConf; market basket analysis; the link between association rules and logical regularities; the anti-monotonicity of support; the APriori algorithm: breadth-first search for frequent itemsets and recursive extraction of rules; modifications: search on a random subsample, item hierarchies, incremental and decremental algorithms; the FP-growth algorithm: the prefix FP-tree, building it in one pass over the transaction database, computing the support of any itemset from the tree, the conditional FP-tree and recursive search for frequent itemsets; the speed of FP-growth compared with APriori.

Collaborative filtering problems: transaction data, the clients–items matrix, filling in missing values and ranking recommendations, the Netflix Prize competition; user-based and item-based correlation methods, the Pearson, cosine and Jaccard similarity measures, Fisher's exact test; the sparse linear model SLIM; latent models and matrix factorisations: SVD, the alternating least squares method ALS, the latent factor model LFM and stochastic gradient, nonnegative matrix factorisations NNMF, PLSA; taking implicit preferences into account (implicit feedback); matrix factorisation as a linear autoencoder, neural network and denoising autoencoders; taking features of clients and items into account, factorisation machines libFM, a hypergraph topic model; the quality of recommendations: RMSE, precision@k, recall@k, diversity, novelty, coverage, serendipity, bias, offline and online evaluation.

The learning-to-rank problem: recovering a partial order relation, the pointwise, pairwise and listwise approaches, the linear ranking model, applications; ordinal regression and the ordinal SVM (SVOR); the pairwise criterion, Ranking SVM, RankNet, the transition to the listwise method LambdaRank; ranking search results: query–document objects, assessor ratings, text, link and click features, TF-IDF and its probabilistic meaning, Okapi BM25, PageRank; quality criteria: precision and recall, MAP, the share of defective pairs and its link with AUC, DCG and NDCG, the pFound user behaviour model; the DSSM neural model: a siamese network, word hashing by letter trigrams, training on clicks with negative sampling.

Probabilistic topic modelling of text document collections: the bag-of-words and conditional independence hypotheses, the maximum likelihood criterion, three interpretations of the problem (soft clustering, stochastic matrix factorisation, autoencoder); the lemma on maximising a function on unit simplices; the ill-posedness of the problem and the additive regularisation of topic models ARTM, the EM algorithm and its interpretation; PLSA and LDA, the Dirichlet distribution, the maximum a posteriori probability and giving up Bayesian inference; smoothing, sparsing, background topics, decorrelation; multimodal and multilingual models, unigrams and bigrams; regularisers for classification and regression on texts, for word co-occurrence in short texts, for links between documents and for topic hierarchies.

Sequence-to-sequence problems: the recurrent network, the encoder–decoder architecture and its limitations; the recurrent network with an attention model, interpreting the attention matrix in machine translation and image captioning; similarity functions, the query–key–value attention model, the attention formula, self-attention; multi-head attention and its regularisation, hierarchical attention for document classification, the graph attention model GAT; the transformer for machine translation: the encoder, positional encoding, layer normalisation and skip connections, the decoder with masking, the training criterion, BLEU and WER; the transformer encoder BERT, the MLM and NSP criteria, fine-tuning, BPE and WordPiece tokenisation, model sizes.