Neighborhood Aggregation Embedding Model for Link Prediction in Knowledge Graphs

Abstract

Link prediction has become a hot topic of knowledge graphs (KGs) in recent years. It aims at predicting missing links between entities to complement KGs. The most successful methods for this problem are embedding-based. Most previous works only consider the triples to learn the embeddings of entities and relations, so the information they can utilize is limited. However, KGs are graph-structured data, we can use the neighborhood information to improve the quality of embeddings, thus improving the performance of link prediction task. In this paper, we propose NAE (neighborhood aggregation embedding model), a novel approach for link prediction. NAE consists of an aggregator and a predictor. The aggregator aggregates the embeddings of multi-order neighbors with different weights to generate a new embedding for each entity. Further analysis shows that the performance of some existing methods such as TransE and DistMult can be improved by integrating our aggregators. The predictor predicts the probability distributions of target entities. It uses convolutional neural network (CNN) to capture more interactions between the new entity embeddings and the relation embeddings. We also propose a highly parameter efficient model NAE-S by simplifying the predictor, which can obtain competitive performance with fewer parameters. Compared with DistMult, NAE-S achieves the same performance with 16x fewer parameters. Experimental results show that our method outperforms several state-of-the-art methods on benchmark datasets.