Distribution Shift

Distribution shift occurs when the data a model encounters in deployment differs from its training data, through changes in inputs, label frequencies, relationships, domains, or time. Researchers study detection and measurement with density-ratio methods, statistical tests, and representation-based diagnostics, and seek to reduce failures through domain generalization, importance weighting, calibration, and test-time adaptation. Evaluation spans settings such as medical imaging across hospitals, temporal forecasting, and vision under changes in weather or sensors, where random splits can conceal deployment risks. A central tension is that adaptation can improve performance but also amplify errors when shift estimates are unreliable or target labels are scarce, making dependable monitoring and realistic evaluation essential.