Multi-view entity attribute prediction model based on entity association and temporal dependency

Zhengzuo Li, Chengxi Piao, Dianhui Chu, Zhiying Tu, Xin Hu, Deqiong Ding

Abstract: Temporal knowledge graph completion (TKGC) is a significant research problem in the field of temporal knowledge graphs (TKGs). Existing studies primarily focus on inferring missing relationships between entities, often overlooking the rich and valuable entity attributes. As part of graph quality assessment, predicting and completing entity attributes is also a crucial aspect of TKGC tasks, playing a vital role in various embedding learning and prediction applications. However, predicting entity attributes remains challenging due to the complex interactions within the graph, temporal dynamics, and the propagation of changes’ impacts throughout the graph. In this paper, we propose a new model, AT-EAPM, to integrate the association dependencies and dynamic temporal dependencies of entities, attributes, and graph structures in TKGs for entity attribute prediction. The model employs a multi-view extraction module to obtain three sets of embedding information for each time window in the knowledge graph (KG): entity attributes, attribute change influence, and attribute influence paths. Due to the high-dimensional multi-view representations and sequential modeling over historical graph snapshots, our method involves intensive matrix operations that are computationally demanding when deployed at scale. This approach aims to explore the deep association features between entities and attributes. Gated recurrent unit (GRU) is used to encode the association features of combined entity and attribute neighborhood representations in each time window, learning temporal dependency features. Based on entity association and temporal dependencies, the model predicts entity attributes. Experiments on datasets with single-attribute and multi-attribute entities demonstrate that multi-view entity association and temporal dependency features are effective, achieving high accuracy. Our implementation adopts mini-batch training with distributed data parallelism (4 GPUs), reducing training time through GPU acceleration. This demonstrates the potential for leveraging high-performance computing (HPC) infrastructure in real-world scenarios involving large or frequently updated temporal knowledge graphs.

External IDs:dblp:journals/tjs/LiPCTHD25