pipette
ENEnglish

Selective Next-Location Prediction with MetaModelSelect: A Lightweight Post-Hoc Confidence Metamodel

Calvin Jary, Nafiseh Kahani

PreprintUso en el mundo real

En palabras de los autores

Next-location prediction can support proactive mobile-network decisions, but acting on unreliable predictions can waste resources. We apply MetaModelSelect, a lightweight post-hoc confidence metamodel, to selective cellular next-location prediction. The mobility-specific model combines a learned embedding of the predicted location with the four largest softmax probabilities and Softmax Concentration. It learns whether a frozen recurrent neural network's top-1 prediction is correct, using only its outputs and without retraining the backbone. On path-balanced Changchun cellular mobility trajectories, we evaluate backbones trained with cross-entropy (CE) and self-adaptive training plus entropy minimization (SAT+EM), comparing against eight scalar confidence scores, two location-aware lookup methods, and a raw-logit multilayer perceptron. Relative to maximum softmax probability, MetaModelSelect achieves mean per-run reductions in area under the risk--coverage curve (AURC), excess AURC (E-AURC), and area under the generalized risk--coverage curve (AUGRC) of 2.6%, 4.2%, and 3.0% for CE, and 2.0%, 3.3%, and 1.9% for SAT+EM. It also achieves the lowest reported mean risk at every evaluated partial coverage. The results show that MetaModelSelect improves confidence ranking in this mobility setting while leaving the underlying location predictor unchanged.

Resultado principalEl resumen no menciona limitaciones.

Apareció: jueves, 15 de octubre. arXiv. Preprint, todavía sin revisión por pares.

Comentario de los autores: accepted at IEEE International Conference on Tools with Artificial Intelligence (ICTAI) 2026