Predicting rare, high‑cost events in multivariate time‑series systems is challenging because precursors are seldom observed, leaving pure data‑driven models with little signal. Engineers operating these systems, however, possess domain knowledge that precedes such events; we refer to this as a semantic prior. KITE injects the semantic prior into a self‑supervised event forecaster via two switchable mechanisms:
- Covariate mechanism: treats the prior as an additional covariate concatenated with the original features;
- Attention‑bias mechanism: adds a prior‑driven bias term to the attention computation.
Both mechanisms reduce to the identity function when no prior is supplied, so a single architecture can accommodate any prior type without per‑dataset architecture search. Empirical results show that when the prior supplies information absent from the observed window, performance improves markedly. For short contexts, a periodicity‑clock prior raises pooled h‑AUROC by 14% over the no‑prior baseline (p<0.05), and consistent gains are observed across several real‑world datasets. The work demonstrates a systematic way to embed domain knowledge into deep temporal models, offering a new avenue for rare‑event forecasting.
Review: KITE’s design elegantly balances model generality with efficient prior utilization, eliminating the need for cumbersome architecture tuning while providing a unified interface for diverse priors, making it highly practical for real‑world deployments.