Using Pairwise Occurrence Information to Improve Knowledge Graph Completion on Large-Scale Datasets

TL;DR

This paper introduces a method leveraging pairwise occurrence information to enhance knowledge graph completion, especially on large datasets, by improving negative sampling and joint learning, leading to better accuracy with smaller batch sizes.

Contribution

The authors propose a novel approach that uses entity-relation pair occurrences to improve training efficiency and performance of bilinear models on large-scale knowledge graphs.

Findings

Significant performance improvement on three standard datasets.

Outperforms baseline by 2.8% on hits@1 for a dataset with 2 million entities.

Effective especially with smaller batch sizes and fewer negative samples.

Abstract

Bilinear models such as DistMult and ComplEx are effective methods for knowledge graph (KG) completion. However, they require large batch sizes, which becomes a performance bottleneck when training on large scale datasets due to memory constraints. In this paper we use occurrences of entity-relation pairs in the dataset to construct a joint learning model and to increase the quality of sampled negatives during training. We show on three standard datasets that when these two techniques are combined, they give a significant improvement in performance, especially when the batch size and the number of generated negative examples are low relative to the size of the dataset. We then apply our techniques to a dataset containing 2 million entities and demonstrate that our model outperforms the baseline by 2.8% absolute on hits@1.