Mining rules on knowledge graph embeddings : reducing rule set sizes and raising explainability

Abstract

With knowledge graph completion methods we can predict missing links in knowledge graphs. Two prominent sub groups of knowledge graph completion methods are rule-based and embedding-based methods - with Any BURL as an example for the first group and ExpressivE as an example for the second group. Both groups have complemental features:Rule-based methods are explainable and applicable to unseen data, while embedding-based methods are not or less explainable and can not predict links on unseen data.However, the performance of embedding-based methods tends to be better on some tasks.Suppose we could combine rule-based and embedding-based knowledge graph completion methods, we could have the best of both worlds: State-of-the-art performance,explainability and the ability to generalize to unseen data. One approach to add external knowledge into rule-based methods, for example, from embedding-based methods, is by reranking the logical rules or by selecting smaller rule sets based on a decision criterion extracted from an embedding-based method. A particularly promising embedding-based method for extracting rule priorities is ExpressivE, which was designed to learn logical rules.In our thesis, we show that ExpressivE is not only capable of learning logical rules, but indeed learns AnyBURL’s Uc rules by introducing the novel concept of Rule Conformance and by evaluating ExpressivE’s Rule Conformance on the WN18RR dataset. Then, we propose the Filtered-MRR and MRR-Evidence Score, two methods to extract a rule prioritization from ExpressivE. We show that we can drastically reduce the size of rulesets needed to achieve state-of-the-art performance with selection methods based on those scores - both on the WN18RR and the InductiveWN18RR dataset. Therefore, our selection methods are applicable both in the transductive setting, i.e., when each testentity was part of the training set, as well as in the inductive setting, when the testinstances were not part of the training set.With our proposed selection methods, we combine rule-based methods with embedding-based methods, drastically reduce the rule set sizes and thereby increase the explainability of knowledge graph completion methods. Additionally, we use embedding-based methods in the inductive setting, which was previously not possible.