Set out the aims, stages and techniques of training a model

Model training has three stages. Method selection depends on whether the data are labelled, the type of target (numerical for regression, categorical for classification), the type and number of features, and the size of the dataset. Performance evaluation uses the right metrics on data held out from training, with cross-validation. Tuning adjusts hyperparameters, often with a grid search, to balance bias and variance: high error on both training and validation data means underfitting, and a large gap means overfitting. Common remedies for unbalanced classes are undersampling the majority or oversampling the minority, in the training data only. Ceiling analysis locates which part of a pipeline most limits performance. Accuracy alone can mislead.