.png){kind=icon}
.png){kind=icon}
Return
Representation learning on knowledge graph: a path attention-based method
.png){kind=icon}
PRE
AI
DOI：10.1007/s11276-025-04041-y.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge representation learning maps entities and relations in the knowledge graph to a continuous and dense space for feature representation. With the development of time, the amount of data contained in the knowledge graph gradually increases, and the spatial structure within the neighborhood of the entity becomes increasingly complex. The existing translation models only target the knowledge triple structure itself, failing to take into account the structural relationships between entities and lacking consideration of knowledge completeness. The knowledge representation learning algorithm based on convolutional neural networks has the problem of knowledge overlap when capturing information from the neighborhood of the entity, and the distinguishability of features is affected. To address these limitations, this paper proposes a knowledge representation learning method based on relational path attention. The model consists of two modules: the encoder and the decoder. Firstly, the neighborhood space of the target entity is defined as two parts: (1) the star knowledge structure formed by the target entity and the first-order neighbor entities; (2) the multi-branch chain structure composed of multi-order relational paths starting from the first-order neighbor entities. Then, the relation fusion algorithm is used to process the chain structure to obtain the first-order message source that aggregates the information of the multi-order neighbor entities. For the star-shaped knowledge structure, to prevent the structural information from being destroyed during the representation learning process, the attention mechanism is used to aggregate the message source information to achieve the vector representation of the target entity and relations. We perform the link prediction task on the FB15K-237 and WN18 datasets for the proposed model and the baseline model, respectively. The experimental results show that the proposed model has a significant performance improvement compared to the existing baseline models, demonstrating the effectiveness and superiority of the algorithm.
Keywords:
Knowledge graph
Knowledge representation learning
Graph attention network
Relation paths
Hierarchical structures
Journal
IF:
2.1
Papers:
308
Citations:
5.4K