Set Prediction

Set prediction studies how to produce a collection of outputs when the number, identity, or exact answer is uncertain, rather than forcing one fixed prediction. Common approaches use learned queries with Hungarian matching, set transformers, or multi-Bernoulli models for structured outputs, while conformal methods calibrate prediction sets to provide coverage guarantees. These methods support tasks such as object detection, segmentation, event forecasting, and scientific reconstruction, as well as uncertainty-aware classification and decisions. A central challenge is balancing coverage and completeness against false positives, set size, computational cost, and reliability when data distributions shift.