Supervised learning trains a model from data paired with labels. The model learns to predict those labels for new records, documents, images, measurements, or other inputs. section 1.2
The predictor’s role depends on the label type:
- Classification predicts categorical labels, such as whether an image shows a cat, dog, or hamster. section 1.2
- Regression predicts numeric labels, such as a numeric quality measure or the number of distinct animals in an image. section 1.2
For outlier detection, supervision becomes practical after enough flagged records have been assessed and labeled. Classifiers can then distinguish points of interest from ordinary records, use boosted models or AutoML tools, and sometimes provide more interpretable models than dedicated outlier detectors. section 4.3.3
From detector feedback to classification
A long-running detector flags records, and people label them as points of interest or not. With sufficient labeled examples, a classifier can replace or supplement the detector. A shallow decision tree may also be converted into rules for an existing rules-based system. section 4.3.3
