Generative Diffusion Prior Distillation for Time-Series Classification

A new arXiv paper (2605.11414) introduces Generative Diffusion Prior Distillation (GDPD) to address the challenge of classifying time-series data when only partial prefixes are available due to latency and cost constraints. Traditional classifiers assume full sequences at inference, but partial data often lacks class-discriminative patterns, hindering generalization. The authors propose using knowledge distillation (KD) from a full-sequence teacher model to a partial-sequence student model. However, direct feature matching fails when the generalization gap stems from data differences (full vs. partial) rather than parameter capacity. GDPD employs a generative diffusion prior to provide progressive, diverse, and collective teacher supervision, enabling the student to learn from the teacher's full-context features more effectively. The method is validated on multiple time-series benchmarks, showing improved classification accuracy on partial inputs.