I85PREP
⌘ K
WBCHSE Applied AI study materials
WBCHSE Applied AI study resources

WBCHSE Applied AI: Introduction to Machine Learning Concepts and Workflow

WBCHSE Applied AI · Work through your exam, one topic at a time.

All guide sections

Introduction to Machine Learning Concepts and Workflow

Domain 6 - Introduction to Machine Learning Concepts and Workflow. This domain tests the full pipeline from problem framing to evaluation, plus accurate use of supervised, unsupervised and reinforcement learning vocabulary.

Core rules and concepts

  • Supervised learning uses labelled data: classification predicts a category, regression predicts a number. Unsupervised learning works on unlabelled data: clustering (k-means), dimensionality reduction (PCA), association rules. Reinforcement learning learns from reward signals by trial and error.
  • Workflow: frame the problem → collect data → clean and engineer features → split into train/validation/test → build a naive baseline → train → tune hyperparameters on validation → evaluate once on test → deploy and monitor.
  • Overfitting means low training error but high test error; the model has memorised noise. Underfitting means high error on both; the model is too simple. Remedies: more data, a simpler model, regularisation, early stopping, cross-validation.
  • Evaluation: the confusion matrix gives TP, TN, FP, FN. Accuracy = (TP+TN)/total; precision = TP/(TP+FP); recall = TP/(TP+FN); F1 is their harmonic mean. Regression uses MAE, MSE, RMSE and R².

Worked example

A fraud dataset has 1,000 transactions, 20 of them fraudulent. A model that always predicts not fraud scores 98% accuracy, yet its recall on the fraud class is 0 and it catches nothing. Correct practice: report precision, recall and F1 for the fraud class, use a stratified split, and treat class weighting or resampling as options. Compare every result against the majority-class baseline, which already scores 98%.

Traps that cost marks

  • Evaluating on the training data and reporting that score as performance.
  • Quoting accuracy on an imbalanced dataset where the majority class dominates.
  • Tuning hyperparameters on the test set; tune on validation, test once.
  • Confusing precision (of the items flagged, how many are correct) with recall (of the actual positives, how many were caught).
  • Calling clustering supervised because the analyst knows the expected groups.