Prediction, not explanation
Machine learning
Thirteen tasks, each fitted alongside the simpler method it claims to beat - and told plainly when the simpler one wins.
Machine learning in a thesis usually goes wrong in one of two ways: a model that scores brilliantly because a feature secretly contains the answer, or a headline accuracy that is worse than always guessing the commonest class. ZStat checks for both before it reports anything.
What you can do
Predict a category or a number, with the model chosen by comparison rather than fashion, and the honest cross-validated score reported instead of the best one seen.
Clustering when you do not know the groups in advance, and anomaly detection for cases that do not belong.
Feature selection and hyperparameter search, both run inside the resampling loop - which is the difference between a real estimate and one inflated by having looked.
Forecasting, LSTM and transformer models for series, and text classification.
When labelling is expensive, semi-supervised methods use the unlabelled cases - and are compared against the model trained on the labels alone, which they frequently fail to beat.
What it refuses
- A model built on a feature that encodes the outcome. It would score perfectly in your thesis and fail completely in use.
- An accuracy figure without the majority-class rate beside it. If 85% of your cases are one class, 85% accuracy means the model learned nothing.
- A forecast validated on a shuffled time series, which lets the model train on Wednesday to predict Tuesday.