Construção automática de grafo de conhecimento no domínio do e-commerce

Abstract

Extracting knowledge efficiently, when large volumes of data are generated daily, is still a challenge. In most cases, these data are unstructured, that is, they are presented in textual or visual format without a clear delimitation of the information they contain and the relationships between this information. Thus, as important as correctly extracting knowledge is to represent it and store it so that it is useful. One of the ways to represent (store) this knowledge is through knowledge graphs. These structures represent semantic relationships (edges) between entities (vertices), as the semantic relationship is_a between the apple and fruit entities represented by the triple: is_a(apple,fruit). Thus, this work addresses the automatic construction of a knowledge graph for the e-commerce domain, where the vertices of this graph represent products and characteristics, while the edges connecting these vertices are used to describe the relationship between them. Among the challenges that this work faced is having to deal with unstructured, noisy and incomplete data generated by users in the e-commerce domain. Added to this fact are the semantic challenges of the domain, since e-commerce data carry more semantic value because they are real entities that came from very varied categories and contexts. In order to advance in the investigation of methods to deal with such challenges and peculiarities of the e-commerce domain, in this work two graph models were trained for product recommendation: one of them following distributive approach through the RedisGraph tool, and another that explores latent properties of the distributed methods of knowledge graph embeddings. The results show that the latter can contribute to tasks in the e-commerce domain that aim at product diversity.