Abstract
Temporal knowledge graph (TKG) reasoning aims to predict missing facts based on a given history. Most of the existing methods unifiedly model the evolution process of different events and ignore their inherent asynchronous characteristics, resulting in suboptimal performance. To tackle this challenge, we propose a Natural Evolution-based Dual-level Aggregation framework (NEDA) for TKG reasoning. Specifically, we design a natural division strategy to group TKGs into different patches according to the occurrence of a given target entity. Then, we present a dual-level aggregation scheme to extract local representations from information within patches and then aggregate these representations with adaptive weights as the final entity representations. By assigning varying weights to different patches, this aggregation scheme can incorporate the asynchronous characteristics of event evolution for representation computation, thus enhancing prediction performance. Extensive experiments demonstrate the significant improvement of our proposed model.
Introduction
Temporal Knowledge Graphs (TKGs) represent facts as quadruples $(s, r, o, t)$, where relationships between entities evolve over time. Reasoning on TKGs, specifically extrapolation, involves predicting future facts based on historical data. A significant limitation in current state-of-the-art models is the assumption of unified modeling—treating all historical snapshots as equally relevant or following a linear progression. In reality, events are asynchronous; some have long life cycles, while others are intermittent or irrelevant to specific queries. Standard recurrent or convolutional architectures often struggle to capture these irregular dependencies, as critical signals can be obscured by historical noise.
Methodology: The NEDA Framework
The NEDA framework addresses asynchrony through a three-stage process: Patch Construction, Intra-patch Encoding, and Inter-patch Attention.
1. Patch Construction
Instead of a uniform temporal stream, NEDA segments history into Explicit Patches and Implicit Patches. Explicit patches consist of consecutive snapshots where the target entity is active. Implicit patches cover periods of inactivity for the target entity, during which its neighbors may still evolve. This "natural division" ensures that related events are grouped into cohesive segments.
2. Intra-patch Encoder (Local Level)
Within each patch, NEDA employs a relation-aware Graph Neural Network (GNN) to capture structural dependencies. A temporal aggregator then models the evolution of entity representations within that specific time block, generating local embeddings.
3. Inter-patch Attention (Global Level)
The core innovation lies in the inter-patch aggregation. Treating patches as a sequence, NEDA uses a cross-attention mechanism to assign adaptive weights to different historical segments. This allows the model to "jump" across time, focusing on highly relevant historical "chapters" while ignoring irrelevant noise, effectively handling events with long life cycles.
Experiments and Results
NEDA was evaluated on six datasets, including standard benchmarks (ICEWS14, ICEWS18, ICEWS05-15, GDELT) and two newly introduced datasets (ICEWS22, ICEWS23) curated for more recent temporal data.
Performance and Data Efficiency
NEDA achieved state-of-the-art results across all datasets. On ICEWS benchmarks, it improved Mean Reciprocal Rank (MRR) by an average of 5.6% and Hits@1 by 8.9% compared to leading baselines like TiRGN.
Notably, NEDA demonstrates high data efficiency. As shown in the training data proportion analysis, NEDA outperforms the leading baseline even when trained on only 10% of the data. This suggests that the patch-based logic effectively captures the underlying evolutionary patterns more robustly than standard flat recurrence.
Conclusion
The NEDA framework introduces a novel approach to TKG reasoning by explicitly modeling the asynchronous nature of event evolution. Through hierarchical aggregation and natural patch division, it overcomes the limitations of unified modeling. While the model currently faces challenges regarding computational overhead due to its entity-centric strategy, its superior performance and data efficiency mark a significant advancement in temporal reasoning. Future work will focus on optimizing inference speed and refining multi-topic modeling within patches.
Summary updated