지도 학습

Each training example tells the algorithm what output is desired for an input. A loss function converts prediction errors into a quantity the training procedure can optimize. The choice of loss shapes learning; the metric used to judge the final workflow may be different. Labels can come from measurements, later outcomes, or annotation. Examine disagreements and ambiguous cases rather than assuming every recorded answer is correct. If the label captures an old decision process, the model can reproduce that process’s limitations. Split the data to match how the model will encounter new cases. Random row splits can leak information when repeated records describe the same subject. Forecasts generally need time-respecting evaluation. Fit preprocessing steps only on the training partition before applying them to validation and test examples. After training, inspect performance for relevant classes and operating conditions. Class imbalance can make overall accuracy misleading. Decide how uncertain or unfamiliar inputs should be handled, and retain a route for correcting labels and reviewing systematic mistakes.