Computational Intelligence Techniques for Big Data Analytics

Abstract

The graph-based representations of knowledge can provide extended opportunities for data exploration and re-interpretation of existing data. The study embraced an approach for clustering entities based on structural similarity within a knowledge graph. In this manner, entities are grouped by matching their linked-based structure rather than relational attributes. As an exemplar of graph-based knowledgebases, a subset of Freebase data was utilised. Specifically, the dataset was used to construct a large-scale knowledge graph covering four categories of entities including Science and Technology, Society, Sports, and Time and Space. Using the Louvain algorithm for graph clustering, the study suggested 19 clusters of entities. The computed clusters were examined to be of good coherence based on the measures of graph modularity and cluster density. Furthermore, the clusters are claimed to be interpretable and revealing relationships among entities. On one hand, most clusters represent homogeneous partitions of entities belonging to a specific category of Freebase. On the other hand, a few clusters featured a remarkable variety of entities relating to different categories. Thus, the clusters can reveal latent connections among entities that are not explicitly observed. For a future research direction, the computed clusters can be utilised as a basis for topic exploration and discovery.