Data/feature pre-processing
🡢 Data cleaning: missing, inconsistend, and noisy data
🡢 Missing values: univariate vs multivariate features, nearest neighbour imputation
🡢 Feature scaling: standard, min-max, max absolute scaling, uniform/Gaussian distribution mapping
🡢 Feature normalisation
🡢 Feature encoding/embedding
🡢 Discretisation: k-bins, feature binarisation
🡢 Label balancing: up-sampling, down-sampling, advanced balancing methods
Evaluating a model
🡢 Confusion matrix in binary vs multi-class classification
🡢 Classification metrics: micro/macro/weighted averages
🡢 Training and test data splitting: pitfalls, cross-validation, k-fold stratified cross-validation
Ensembles and Neural Networks
🡢 Ensembles: baggin meta-estimator, forests (random forests and estremely randomised trees)
🡢 AdaBoost
🡢 Stacked ensembles
🡢 Multilayer Perceptron
Latent representation and embeddings
🡢 MNIST dataset as benchmarking system
🡢 Use BAE to build simple and boosting-based autoencoders
🡢 Variational Autoencoders (VAEs)
🡢 Vector Quantized Variational Autoencoders (VQ-VAEs)
🡢 Exemplar Autoencoder
🡢 Attention mechanisms and Transformers
Hyperparameter tuning
🡢 Grid search: drawbacks
🡢 Randomised search: drawbacks
🡢 Bayesian optimisation: advantages and drawbacks
Complex algorithms for Electronic Health Record (EHR) data
🡢 Vanilla-LSTM
🡢 T-LSTM
🡢 BERT-like solution pre-trained (PubMedBERT)
