Reproduce “HEPA: A Self-Supervised Horizon-Conditioned Event Predictive Architecture for Time Series” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jonas Petersen, Gian-Alessandro Lombardi, Riccardo Maggioni, Camilla Mazzoleni, Federico Martelli, Philipp Petersen
HEPA learns future representations from unlabeled histories, then freezes its encoder and adapts a predictor to estimate event risk across horizons. Its survival CDF ensures risk cannot decrease with horizon. On C-MAPSS-3, full-label h-AUROC is 0.84 ± 0.01, versus 0.79 ± 0.01 for PatchTST and 0.78 ± 0.02 for MAE; it leads on 10 of 14 benchmarks, but not all.
Key insights
- RUL prediction and anomaly detection share a fundamental task of estimating event probability within a future horizon, allowing for a unified architectural approach.
- A theoretically derived information-theoretic bound (Proposition 1) formalizes how reduced pretraining prediction error correlates with increased event-relevant information retention in the encoder.
- Predictor finetuning, tuning ~11x fewer parameters than end-to-end training, provides comparable performance at full label budgets while being more expressive than a linear probe.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.