A Complex Network Model for Knowledge Graphs' Relationships

When dealing with Knowledge Graphs (KGs) structure, content, and quality, the focus is generally on entities. We show here that modelizing individual relationships, with their evolution, is also possible. This brings new opportunities for conducting various analyses on KGs or for improving benchmarks. Relationships matter: we present KRELM, the first – simple yet powerful – graph generative model able to (i) closely mimic a large set of crowdsourced KG relationships and (ii) simulate well their evolution. In particular, for crowdsourced KGs, we show that the decentralized process of crowdsourcing is able to produce distribution patterns that are reproducible using KRELM. In this model, the facts of a relationship are added one by one, either by adding new entities or by describing existing ones, with asymmetric attachment between subjects and objects. The theoretical analysis of KRELM enables us to understand the fundamental dynamics of a knowledge graph, where the distribution of facts for each relationship follows an exponential law for subjects and a power law for objects. Our experiments show on several major KGs that KRELM perfectly reproduces the structure of a large part of their relationships. Moreover, a longitudinal study of Wikidata underlines our model’s relevance in predicting this structure’s evolution.