Table of Contents
| Classification | Regression | |
|---|---|---|
| Output | discrete label (finite classes) | continuous value |
| Goal | which category? | what numeric value? |
| Loss | cross-entropy, hinge | MSE, MAE |
| Metrics | accuracy, F1, AUC | RMSE, R², MAE |
Measuring confidence
- Classification —
max(softmax)as confidence; entropy of the output (lower → more confident); MC dropout (dropout at inference to estimate uncertainty). - Regression — confidence/prediction intervals (the range the true value likely falls in).
Converting between them
Classification → regression: predict a continuous score + threshold (sigmoid → threshold at 0.5 for binary; regress toward 0/1 targets, argmax for multi-class). Useful when you want uncertainty estimates, regression-based models, or to combine losses (multi-task). Risk: regression may ignore class boundaries → poor accuracy.
Regression → classification: bin the continuous output into buckets (e.g. age → 0–18/19–35/…), turning it into ordinal/multi-class. Useful when exact values are noisy/irrelevant, you care about ranges, or to handle outliers/imbalance. Risk: discretization loses information/resolution, and ordinal ordering can be lost in standard classification.
Related: loss-functions, logistic-regression, evaluation-metrics, class-imbalance