Research map / Graph learning / Knowledge and temporal graphs

DECRL: A Deep Evolutionary Clustering Jointed Temporal Knowledge Graph Representation Learning Approach

NeurIPS 2024 · Graph learning › Knowledge and temporal graphs · approach: Temporal knowledge graph reasoning · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DECRL, a TKG representation learning model that jointly learns deep evolutionary clustering of entities (soft fuzzy c-means with cluster-aware Hungarian alignment and temporal smoothness loss) and an implicit cluster-graph correlation encoder guided by a global graph, evaluated on seven TKG datasets for event prediction.

Problem

Existing TKG representation learning approaches cannot capture the temporal evolution of high-order correlations among entities for event prediction.

Approach

Integrates deep evolutionary clustering into TKG representation learning: soft overlapping fuzzy clusters are aligned across consecutive timestamps via a cluster-aware Hungarian matching and smoothed with a temporal loss, while an implicit correlation encoder models pairwise cluster interactions weighted by a global-graph spectral clustering prior.

Key results

• On ICEWS14, DECRL reaches 62.61 MRR and 93.03 Hits@10, versus 56.15 MRR and 85.89 Hits@10 for the best baseline DHyper.
• On GDELT, DECRL achieves 27.58 MRR versus 24.61 for TiRGN, and on WIKI and YAGO it reaches 99.67 and 99.56 MRR, slightly above DHyper.
• The paper reports average gains of 9.53% MRR and 14.68% Hits@10 over the best baseline in the abstract, while the results section states 13.24% MRR, which is inconsistent.

Measured effects reported in the paper

Surprising findings

• On the entity prediction task on GDELT, DECRL does not beat GTRL in MRR or Hits@1 (37.24 vs 39.23 MRR), though it leads in Hits@3 and Hits@10.

Limitations

• The implicit correlation encoder assumes uniform correlations across all clusters, which the authors note may not reflect real inter-organization interactions.
• Temporal modeling only considers the previous and current timestamps rather than continuous influence from multiple past timestamps.
• Entity prediction results are weaker than SOTA on MRR and Hits@1 on GDELT.

Experimental setup

Datasets: ICEWS14 ICEWS18 ICEWS14C ICEWS18C GDELT WIKI YAGO

Benchmarks: ICEWS14 ICEWS18 ICEWS14C ICEWS18C GDELT WIKI YAGO

Metrics: MRR Hits@1 Hits@3 Hits@10

Baselines: TTransE HyTE RE-NET Glean TeMP RE-GCN DACHA TiRGN TITer EvoExplore GTRL DHyper

Builds on: RE-GCN ConvTransE TiRGN DHyper

Closest papers by meaning

• TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs (NeurIPS 2024 · similarity 0.88)
• Deep Temporal Graph Clustering (ICLR 2024 · similarity 0.86)
• Large Language Models-guided Dynamic Adaptation for Temporal Knowledge Graph Reasoning (NeurIPS 2024 · similarity 0.86)
• Integrating Large Language Models and Möbius Group Transformations for Temporal Knowledge Graph Embedding on the Riemann Sphere (AAAI 2025 · similarity 0.86)
• Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with Large Language Models (NeurIPS 2025 · similarity 0.85)
• HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces (AAAI 2024 · similarity 0.85)