Classification
Classification predicts a discrete label for each input, from spam-or-not to many mutually exclusive categories.
Predicting categories
Classification is supervised learning where the target is a discrete class rather than a number. Binary classification chooses between two labels; multiclass chooses among several; multilabel assigns any subset. Most classifiers output a score or probability per class, then pick the highest.
Common algorithms
- Logistic regression: linear boundary, calibrated probabilities.
- Decision trees and random forests: axis-aligned rules, nonlinear boundaries.
- Gradient boosting: strong tabular performance.
- SVM: maximum-margin boundaries, kernels for nonlinearity.
- k-NN and naive Bayes: simple, fast baselines.
Decision boundaries
A classifier partitions feature space into regions, one per class. The boundary between regions may be linear (logistic regression, linear SVM) or highly nonlinear (trees, kernel SVM, networks). The right complexity depends on the data and the bias-variance tradeoff.
Measuring success
Accuracy alone misleads on imbalanced data. Use precision, recall, and F1, inspect the confusion matrix, and for probabilistic outputs read the ROC and AUC. Choosing a threshold trades false positives against false negatives to match the costs of the application.
In diagnostics, a classifier might sort plasma shots into stable and disruptive categories; there, missing a disruption (a false negative) is far costlier than a false alarm, so the threshold is tuned toward recall.