Da Tong’s research while affiliated with Institute of Microelectronics, Chinese Academy of Sciences and other places

Knowledge graph embedding (KGE) is typically used for link prediction to automatically predict missing links in knowledge graphs. Current KGE models are mainly based on complicated mathematical associations, which are highly expressive but ignore the uniformity behind the classical bilinear translational model TransE, a model that embeds all entities of knowledge graphs in a uniform space, enabling accurate embeddings. This study analyses the uniformity of TransE and proposes a novel KGE model called ConvUs that follows uniformity with expressiveness. Based on the convolution neural network (CNN), ConvUs proposes constraints on convolution filter values and employs a multi-layer, multi-scale CNN architecture with a non-parametric L2 norm-based scoring function for the calculation of triple scores. This addresses potential uniformity-related issues in existing CNN-based KGE models, allowing ConvUs to maintain a uniform embedding space while benefiting from the powerful expressiveness of CNNs. Furthermore, circular convolution is applied to alleviate the potential orderliness contradictions, making ConvUs more suitable for conducting uniform space KGE. Our model outperformed the base model ConvKB and several baselines on the link prediction benchmark WN18RR and FB15k-237, demonstrating strong applicability and generalization and indicating that the uniformity of embedding space with high expressiveness enables more efficient knowledge graph embeddings.

Temporal knowledge graphs play an increasingly prominent role in scenarios such as social networks, finance, and smart cities. As such, research on temporal knowledge graphs continues to deepen. In particular, research on temporal knowledge graph reasoning holds great significance, as it can provide abundant knowledge for downstream tasks such as question answering and recommendation systems. Current reasoning research focuses primarily on interpolation and extrapolation. Extrapolation research aims to predict the likelihood of events occurring in future timestamps. Historical events are crucial for predicting future events. However, existing models struggle to fully capture the evolutionary characteristics of historical knowledge graphs. This paper proposes a multi-scale evolutionary network (MSEN) model that leverages Hierarchical Transfer aware Graph Neural Network (HT-GNN) in a local memory encoder to aggregate rich structural semantics from each timestamp’s knowledge graph. It also utilizes Time Related Graph Neural Network (TR-GNN) in a global memory encoder to model temporal-semantic dependencies of entities across the global knowledge graph, mining global evolutionary patterns. The model integrates information from both encoders to generate entity embeddings for predicting future events. The proposed MSEN model demonstrates strong performance compared to several baselines on typical benchmark datasets. Results show MSEN achieves the highest prediction accuracy.

As an essential method of knowledge graph fusion, entity alignment aims to find entities that refer to the same real-world object in different knowledge graphs. Current entity alignment methods usually adopt extra information such as entity names, attribute triples except for structural information of the knowledge graph. However, due to the difficult availability and possibly low effectiveness of extra information, it is necessary to improve the performance of entity alignment when using only structural information. In this paper, a novel entity alignment method based on the multiscale convolutional graph network (MCEA) is proposed, which utilizes only structural information of the graph for entity alignment. Firstly, the convolution region of long-tail entities is extended to enhance the ability of information capture of the graph network. Secondly, intermediate results from semi-supervised learning are introduced to negative sampling in order to improve sampling quality. Thirdly, the stable marriage algorithm is chosen as the alignment strategy to obtain final alignment results. The experimental results show that this method has achieved better performance on Hits@K and MRR than the state-of-the-art methods, especially in the case of less labeled data. Moreover, we also find that the impact of the alignment strategy has become limited when the model generates sufficient accurate entity embeddings.