Latent Dynamics · 2026-10-12
KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting
Jonas Petersen, Pratik Jawahar, Riccardo Maggioni, Gian-Alessandro Lombardi, Matteo Cozzi, Marek Gajewski, Federico Martelli, Camilla Mazzoleni, Philipp Christian Petersen
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
KITE (Knowledge-Injected Temporal Encoder) injects engineering “semantic priors” into a self-supervised horizon-conditioned event forecaster using two switchable mechanisms: a covariate route and an attention-bias route (each identity at its null). The authors report that informative priors help in short contexts and out-of-domain transfer by providing information the observed window lacks, while redundant priors and the graph bias show no measurable gain.
Why it matters
Rare, high-cost events in multivariate time series have scarce precursors, limiting purely data-driven learning. KITE’s switchable, attributable prior injection offers a single backbone architecture that can accept heterogeneous prior types via inputs rather than per-dataset architecture changes, aiming to support contextual conditioning and scaling toward cross-domain pretrained time-series world models that carry governing structure explicitly.
Method
- Two-stage JEPA-based event forecaster: Stage-1 pretrains a permutation-invariant, cardinality-free encoder with self-supervised latent prediction; Stage-2 freezes the encoder and trains a compact event-probability head.
- Semantic priors enter as switchable inputs via (I) covariate mechanism (static descriptors to encoder, future-known covariates to predictor, and past-only derived residuals), and (II) attention-bias mechanism (graph driver adjacency added to encoder attention logits).
- The authors evaluate h-AUROC for event forecasting and use per-seed paired ablations where the prior is zeroed/absent under the same recipe for “identity at the null” comparisons.
Limitation
The paper scopes evidence to short context and two evaluated prior mechanisms; it also notes that several taxonomy categories remain conceptual, mechanism (II) is bounded but undetermined, and the strongest transfer target is synthetic.
Abstract (from arXiv)
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.
Related papers
- Pythia: Toward Foundation World Models for Multimodal Time Series
- Does Joint-Embedding Predictive Architecture Pretraining Help Time Series Forecasting?
- TAM: Task-Aware Memory Distillation for Efficient Spatiotemporal Prediction
- BeliefGraph-JEPA: Structured Latent World Models for Action-Conditioned Time Series
- CTIA-ODE: Continuous-Time Inter-Variable Attention for Long-Term Time Series Forecasting
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-08 · Indexed by Pub-AI: 2026-10-12 UTC
KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting
Preprint posted: 2026-10-08 · Original record ↗
- arxiv
- 2610.12738
Explicit source versions (1)
- v1 · 2026-10-08
- Preprint available
Source date: 2026-10-08 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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.