Unified Representation Learning Approaches

Throughout this monograph, different representation learning techniques have demonstrated that propositionalization and embeddings represent a multifaceted approach to symbolic or numeric feature construction, respectively. At the core of this similarity between different approaches is their common but implicit use of different similarity functions. In this chapter, we take a step forward by explicitly using similarities between entities to construct the embeddings. We start this chapter with Sect. 6.1, which presents entity embeddings, a general methodology capable of supervised and unsupervised embeddings of different entities, including texts and knowledge graphs. Next, two unified approaches to transforming relational data, PropStar and PropDRM, are presented in Sect. 6.2. These two methods combine propositionalization and embeddings, benefiting from both by capturing relational information through propositionalization and then applying deep neural networks to obtain dense embeddings. The chapter concludes by presenting selected methods implemented in Jupyter Python notebooks in Sect. 6.3.