Hawkes based Representation Learning for Reasoning over Scale-free Community-structured Temporal Knowledge Graphs

TL;DR

HERLN introduces a Hawkes process based evolutional representation learning network for reasoning over scale-free community-structured temporal knowledge graphs. It jointly models static community structure via embedding initialization, dynamic evolution via a Hawkes-weighted relational GCN, and query-specific decoding via a FiLM-modulated ConvTransE to mitigate long-tailed biases. The method yields significant improvements in entity and relation prediction across four benchmarks and is supported by ablations that confirm the contribution of each component. By integrating real-world network properties with temporal decay, HERLN enhances extrapolative reasoning in dynamic KGs with practical implications for event forecasting and knowledge-driven question answering.

Abstract

Temporal knowledge graph (TKG) reasoning has become a hot topic due to its great value in many practical tasks. The key to TKG reasoning is modeling the structural information and evolutional patterns of the TKGs. While great efforts have been devoted to TKG reasoning, the structural and evolutional characteristics of real-world networks have not been considered. In the aspect of structure, real-world networks usually exhibit clear community structure and scale-free (long-tailed distribution) properties. In the aspect of evolution, the impact of an event decays with the time elapsing. In this paper, we propose a novel TKG reasoning model called Hawkes process-based Evolutional Representation Learning Network (HERLN), which learns structural information and evolutional patterns of a TKG simultaneously, considering the characteristics of real-world networks: community structure, scale-free and temporal decaying. First, we find communities in the input TKG to make the encoding get more similar intra-community embeddings. Second, we design a Hawkes process-based relational graph convolutional network to cope with the event impact-decaying phenomenon. Third, we design a conditional decoding method to alleviate biases towards frequent entities caused by long-tailed distribution. Experimental results show that HERLN achieves significant improvements over the state-of-the-art models.

Figures (3)

• Figure 1: Overall framework of the proposed HERLN model. HERLN is consists of three modules: an embedding initializing module, an evolution encoding module and a conditional decoding module. First, the embedding initializing module extracts the community structural information in input TKGs to get the initialized embeddings. Then the evolution encoding module updates the embeddings with the candidate-related historical structure to learn the evolutional patterns of events. Finally, the conditional decoding module reasons according to the embeddings and gets scores of the candidate entities, then select the entity with the highest score as the results. The input TKG contains 6 nodes (marked as A to F), three different types of edges (indicated in red, blue and green respectively) and three timestamps (from $t_1$ to $t_3$). Additionally, the model receives an incomplete quadruple as query. The query given in the figure is $(A,r_3,?,t_3)$.
• Figure 2: The update process of HRGCN. (a) is a historical structure of node A, (b) shows the update process of a traditional RGCN, which does not use timestamp information and (c) shows the update process of HRGCN, which takes the time interval as decaying weight to represent the event declines over time.
• Figure 3: Time cost analysis on ICEWS14s and YAGO.