SOCP method slashes regional coverage gaps by 7.1% without retraining models

A new paper from Louis Berthier, Ahmed Shokry, and colleagues presents Self-Organized Conformal Prediction (SOCP), a calibration method that tackles a long-standing problem in conformal prediction: regional undercoverage. Standard conformal prediction guarantees marginal coverage (e.g., 90% overall), but safety-critical subgroups may have much lower coverage. SOCP uses a Self-Organizing Map (SOM) to discover natural groups in the input space without any supervised labels. At test time, it draws a local calibration buffer from the query's best-matching unit (BMU) cell or a fixed grid neighborhood, producing group-local validity. The method works for regression and classification, across tabular features and image embeddings, and does not require retraining the predictor or changing the nonconformity score.

On eight benchmarks, SOCP reduces the weighted regional coverage gap on 7/8 datasets, with a mean paired change of -7.1% for a mean prediction-set size increase of only 6.2%. A split-routed variant recovers fixed retrieved-set validity. The approach is computationally light: overhead is negligible on the largest six datasets. By learning groups directly from input geometry, SOCP provides exact validity for BMU-cell retrieval and approximate guarantees for neighborhood buffers — all without supervised partitions or model retraining. This makes it attractive for applications in medical imaging, autonomous driving, and other safety-critical domains where subgroup coverage matters.