Knowledge guides time-series models to forecast rare costly events
KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting
Machine LearningArtificial Intelligence
Summary
Forecasting rare but important events in complex time-based data is very challenging because there are usually very few examples to learn from. The authors show that adding expert knowledge about event causes, which they call a semantic prior, helps a prediction model perform better. They designed KITE, a system that can switch how it uses that knowledge depending on the type of information available, improving accuracy especially when there is little data or when applying to new settings. This approach lets models use human insights alongside data to better predict uncommon but critical events.
What this means in practice
- •For industrial system engineers: Improve forecasting of rare, costly machine failures by injecting domain knowledge into event prediction models.
- •For financial risk teams: Enhance detection of rare adverse market events by using expert insights to supplement scarce labeled examples.
Authors
Jonas Petersen, Pratik Jawahar, Riccardo Maggioni, Gian-Alessandro Lombardi, Matteo Cozzi, Marek Gajewski, Federico Martelli, Camilla Mazzoleni, Philipp Christian Petersen
Abstract
Forecasting rare, high-cost events in multivariate time-series systems is hard because their precursors are seldom observed, so a purely data-driven model has little to learn from. The engineers who run these systems, however, hold knowledge that precedes such events, information absent from the scarce labels. We call this a semantic prior and inject it into a self-supervised event forecaster through two switchable mechanisms, a covariate mechanism and an attention-bias mechanism, each the identity at its null, so one architecture serves every prior type with no per-dataset architecture search. A prior helps when it supplies information the observed window does not already contain: in a short context an informative prior beats no prior in-domain (pooled h-AUROC over the no-prior baseline: a periodicity clock +14%, p<0.001; a conservation residual +47% at a 0.25 label fraction), and a working prior transfers out-of-domain (+53%), substituting for feature transfer, while priors the window already recovers stay inert and the attention-bias mechanism shows no measurable gain. KITE is a switchable, attributable route to contextual conditioning, and a substrate for scaling toward cross-domain, pretrained time-series foundation world models that carry their governing structure explicitly.