Supervised problem types, by the shape of their output Nine glyphs in three families. Classification: binary picks one of two slots, multi-class one of five, multi-label several of five. Numeric: regression places a point anywhere on a continuous line, ordinal regression places it in one of five ordered buckets, forecasting continues a line past the present into a dashed future. Structured: ranking orders items by descending bar length, sequence labeling tags every position in a row, and structured prediction emits a tree whose parts depend on each other. What shape is the answer? Each family differs in what the model is asked to emit — not in how it is trained. CLASSIFICATION NUMERIC STRUCTURED Binaryone of two Multi-classone of many Multi-labelmany of many Regressiona point on a continuum Ordinal regressionbuckets that have an order now Forecastingthe line continues Rankingorder is the answer Sequence labelingone label per position Structured predictionparts constrain each other
Nine supervised problem types, drawn by what the model must emit. The training may look alike; the shape of the answer is what differs.