Unsupervised Relational Representation Learning via Clustering: Preliminary Results

The goal of unsupervised representation learning methods is to learn a new representation of the original data, such that it makes a certain classification task easier to solve. Since their introduction in late 2000s, these methods initiated a revolution within machine learning. In this paper we present an unsupervised representation learning method for relational data. The proposed approach uses a clustering procedure to learn a new representation. Moreover, we introduce an adaptive clustering method, capable of addressing multiple interpretations of similarity in relational context. Finally, we experimentally evaluate the proposed approach. The preliminary results show the promise of the approach, as the models learned on the new representation often achieve better performance and are less complex than the ones learned on the original data representation.