Application of concepts of neighbours to knowledge graph completion1

Abstract

The open nature of Knowledge Graphs (KG) often implies that they are incomplete. Knowledge graph completion (a.k.a. link prediction) consists in inferring new relationships between the entities of a KG based on existing relationships. Most existing approaches rely on the learning of latent feature vectors for the encoding of entities and relations. In general however, latent features cannot be easily interpreted. Rule-based approaches offer interpretability but a distinct ruleset must be learned for each relation. In both latent- and rule-based approaches, the training phase has to be run again when the KG is updated. We propose a new approach that does not need a training phase, and that can provide interpretable explanations for each inference. It relies on the computation of Concepts of Nearest Neighbours (C-NN) to identify clusters of similar entities based on common graph patterns. Different rules are then derived from those graph patterns, and combined to predict new relationships. We evaluate our approach on standard benchmarks for link prediction, where it gets competitive performance compared to existing approaches.

SAFRAN: An interpretable, rule-based link prediction method outperforming embedding models

“…We compare our results with the established latent models RESCAL [Nickel et al, 2011], TransE [Bordes et al, 2013], DistMult [Yang et al, 2015], ComplEx [Trouillon et al, 2016], ConvE [Dettmers et al, 2018], RotatE [Sun et al, 2019], TuckER [Balazevic et al, 2019] and the rule-based methods AMIE+ [Galárraga et al, 2015], RuleN [Meilicke et al, 2018], C-NN [Ferré, 2020], Neural LP [Yang et al, 2017], RLvLR [Omran et al, 2018], DRUM [Sadeghian et al, 2019], GPFL [Gu et al, 2020b]. For details on these approaches we refer to their respective original papers.…”

“…However as the overall dataset is considerably small, the absolute size of its validation set is only 3,134 triples. As rules generated by AnyBURL can generalize to unseen entities [Ferré, 2020], this size may not be sufficient to generate the most optimal clustering. FB15K-237 WN18RR YAGO3-10 Approach MRR hits@1 hits@10 MRR hits@1 hits@10 MRR hits@1 hits@10 Latent Table 7: MRR, Hits@1, Hits@10 results for FB15K-237, WN18RR and YAGO3-10.…”

“…† Results were evaluated with top policy for dealing with same score entities (see Appendix A) and are not directly comparable to other approaches. Results marked with are from [Ruffinelli et al, 2020], ¶ from [Rossi et al, 2021], from [Ferré, 2020], ‡ from [Sadeghian et al, 2019], ♦ from [Gu et al, 2020b], ♣ from [Meilicke et al, 2019], from [Omran et al, 2018] and from [Zhang et al, 2020].…”

“…In [Ferré, 2020] another interesting variation of Noisy-OR, the Dempster-Shafer score [Denoeux, 1995], has been applied. This approach is not based on rules, but relies on another symbolic representation using concepts of nearest neighbours.…”

“…This approach is not based on rules, but relies on another symbolic representation using concepts of nearest neighbours. In their experiments the authors found that the aggregation using the Dempster-Shafer score performs at best 1.3 % worse than the maximum score [Ferré, 2020].…”