Research On Knowledge Graph Embedding Based On Group Feature Interaction

The knowledge graph is a directed graph composed of entities and relationships.The counters in the graph represent various objects in the real world,and the edges in the graph represent the relationships between these objects.Knowledge graphs are usually incomplete,there will be a large number of missing relationships.It will take a lot of manpower and time to complete these missing relationships manually.Therefore,knowledge representation learning is proposed,by embedding entities into a continuous low-dimensional vector space,and then automatically completing the knowledge graph.There are now many knowledge embedding methods,which can be divided into two major categories.The first categories is the Trans-based methods based on translation that starts with Trans E,the second type is a multiplication-based embedding method that starts with RESCAL.From the perspective of the positive interaction between head and tail entities,the existing method can be divided into head and tail entities with only bit-base feature interactions,such as Trans E,Trans H,Trans At,Dist Mult,Complx,etc.,and the other is The overall feature interactions between head and tail entities such as Trans R,CTrans R,RESCAL,etc.Only the model parameters of bit-base feature interactions can be replaced for training,but there are few feature interactions and weak expression ability.For the shortcomings of the above two types of methods,the feature grouping idea is proposed here.The features in the group interact with each other between the head and tail entities,so that the model has better expression ability.At the same time,when the group size is small,the model parameters and training time are similar to bit-base feature interaction models.Based on the idea of feature group interaction,this paper proposes a translationbased model Group E+ and a repetitive-based model Group E* and fully explain the theoretical advantages of the two models.Furthermore,the link prediction experiment results on commonly used public data sets show the performance advantages of the two models proposed,and thus prove the effectiveness of the feature grouping interaction idea.