Evolving Beyond Snapshots: Harmonizing Structure and Sequence via Entity State Tuning for Temporal Knowledge Graph Forecasting

TL;DR

This work tackles the fundamental limitation of stateless temporal knowledge graph forecasting, where entity representations are recomputed anew at each timestamp, causing rapid loss of long-term dependencies. The authors propose Entity State Tuning (EST), an encoder-agnostic framework that maintains persistent entity states via a global memory buffer and a closed-loop evolution, effectively bridging structural and sequential dynamics. EST comprises a Topology-Aware State Perceiver, a Unified Temporal Context Encoder, and a De-confounded State Evolution mechanism with fast/slow memories and counterfactual consistency learning, and it demonstrates state-of-the-art results across four benchmarks with backbone-agnostic gains. The findings highlight state persistence as crucial for robust, long-horizon TKG forecasting and point to future work addressing memory costs and handling unseen entities in dynamic real-world graphs.

Abstract

Temporal knowledge graph (TKG) forecasting requires predicting future facts by jointly modeling structural dependencies within each snapshot and temporal evolution across snapshots. However, most existing methods are stateless: they recompute entity representations at each timestamp from a limited query window, leading to episodic amnesia and rapid decay of long-term dependencies. To address this limitation, we propose Entity State Tuning (EST), an encoder-agnostic framework that endows TKG forecasters with persistent and continuously evolving entity states. EST maintains a global state buffer and progressively aligns structural evidence with sequential signals via a closed-loop design. Specifically, a topology-aware state perceiver first injects entity-state priors into structural encoding. Then, a unified temporal context module aggregates the state-enhanced events with a pluggable sequence backbone. Subsequently, a dual-track evolution mechanism writes the updated context back to the global entity state memory, balancing plasticity against stability. Experiments on multiple benchmarks show that EST consistently improves diverse backbones and achieves state-of-the-art performance, highlighting the importance of state persistence for long-horizon TKG forecasting. The code is published at https://github.com/yuanwuyuan9/Evolving-Beyond-Snapshots

Figures (7)

• Figure 1: From Stateless Snapshot Encoding to Stateful Entity Reasoning in Temporal Knowledge Graphs.
• Figure 2: Efficiency comparison among different sequence encoders on the ICEWS14 dataset.
• Figure 3: Hyperparameter analysis for EST-Mamba (top) and EST-Transformer (bottom).
• Figure 4: Top-5 entities exhibiting the largest state displacement on the ICEWS14 dataset.
• Figure 5: Top-5 entities exhibiting the largest state displacement on the ICEWS18 dataset.