Machine Learning Fundamentals
What this domain tests
Machine Learning Fundamentals tests supervised versus unsupervised learning, regression versus classification, clustering, and model evaluation. Confusion-matrix calculations and overfitting are common exam items.
Core rules and concepts
- Supervised learning uses labelled data. Regression predicts a continuous target, such as price, with metrics MSE, RMSE, MAE and R-squared. Classification predicts a category, such as spam/not spam.
- Unsupervised learning uses unlabelled data. K-means clustering groups similar points; PCA reduces dimensions.
- Split data into training and test sets; use validation or cross-validation for tuning. Overfitting means low training error but poor test performance. Underfitting means poor performance on both.
- Classification metrics: accuracy, precision = TP/(TP+FP), recall = TP/(TP+FN), F1 = 2PR/(P+R), and a confusion matrix.
Worked example
A classifier is tested on 100 cases: TP = 30, FN = 10, FP = 12, TN = 48. Accuracy = (30+48)/100 = 0.78. Precision = 30/(30+12) = 0.714. Recall = 30/(30+10) = 0.75. F1 = 2 x 0.714 x 0.75 / (0.714+0.75) = 0.732. Reasoning: accuracy hides the 12 false positives; precision and recall show different error costs.
Traps that cost marks
Accuracy is misleading on imbalanced data. Do not scale test data using statistics from the test set; fit scaling on training data only. Data leakage occurs if test information enters training. Clustering has no ground-truth labels, so accuracy is not the usual metric. More features do not always improve a model.