Short Note: Supervised vs. Unsupervised Learning

Both are machine-learning paradigms used in data mining; the fundamental difference is whether the training data carry class labels.

1. Supervised learning

Learning from a set of labelled examples - each training record has both the input attributes and the known output (class/target). The learner builds a model (a function $f: X \rightarrow Y$) that predicts the label of unseen records. The problem is described as teaching the machine with a teacher.

• Algorithms: decision trees (ID3, C4.5, CART), naive Bayes, logistic regression, support vector machines, k-nearest neighbour, neural networks, rule-based learners.
• Evaluation: a labelled test set gives accuracy, precision, recall, F1-score, ROC-AUC.
• Applications: spam detection, credit/loan approval, disease diagnosis, price prediction, image classification.
• Cost: collecting labelled data is expensive and time-consuming.

2. Unsupervised learning

Learning from unlabelled data - only the input records are given, and the algorithm must discover the hidden structure, groups or regularities by itself. There is no teacher and no single correct answer.

• Algorithms: clustering (k-means, k-medoids, hierarchical, DBSCAN), association-rule mining (Apriori, FP-growth), dimensionality reduction (PCA), anomaly/novelty detection.
• Evaluation: only internal criteria - cohesion, separation, silhouette coefficient, purity, support/confidence - because there is no ground truth.
• Applications: market-basket segmentation, customer segmentation, document grouping, fraud/anomaly detection, image compression.
• Advantage: no labelling cost; can find patterns no one had thought to look for.

3. Comparison table

Criterion	Supervised learning	Unsupervised learning
Training data	Labelled (input-output pairs)	Unlabelled (input only)
Teacher / feedback	Yes - correct answers are given	No - feedback is absent
Goal	Predict the class/value of new data	Discover hidden structure, groups, rules
Output	Predictive model (classifier/regressor)	Clusters, associations, projections
Evaluation	Accuracy, precision, recall, F1	Internal measures (silhouette, support)
Human effort / cost	High (data labelling)	Low
Complexity of problem	Well-defined target	Open-ended, exploratory
Typical errors	Over-fitting / under-fitting	Choosing k, distance measure, thresholds
Examples	Decision tree, SVM, k-NN, Naive Bayes	k-means, DBSCAN, Apriori, PCA

4. When is which used?

• Use supervised learning when labelled historical examples exist and the goal is prediction or decision support.
• Use unsupervised learning when no labels exist or when the aim is exploration (segmentation, profiling, anomaly hunting).
• Semi-supervised learning combines a small labelled set with a large unlabelled set and is very practical when labelling is costly; reinforcement learning, where an agent learns by reward and penalty from the environment, is a third distinct paradigm.