What Affects the Performance of Models? Sensitivity Analysis of Knowledge Graph Embedding

Knowledge graph (KG) embedding aims to embed entities and relations into a low-dimensional vector space, which has been an active research topic for knowledge base completion (KGC). Recent researchers improve existing models in terms of knowledge representation space, scoring function, encoding method, etc., have achieved progressive improvements. However, the theoretical mechanism behind them has always been ignored. There are few works on sensitivity analysis of embedded models, which is extremely challenging. The diversity of KGE models makes it difficult to consider them uniformly and compare them fairly. In this paper, we first study the internal connections and mutual transformation methods of different KGE models from the generic group perspective, and further propose a unified KGE learning framework. Then, we conduct an in-depth sensitivity analysis on the factors that affect the objective of embedding learning. Specifically, in addition to the impact of the embedding algorithm itself, this article also considers the structural features of the dataset and the strategies of the training method. After a comprehensive experiment and analysis, we can conclude that the Head-to-Tail rate of datasets, the definition of model metric function, the number of negative samples and the selection of regularization methods have a greater impact on the final performance.