Embedding Approaches for Relational Data

​Embedding methods for searching latent representations of the data are very important tools for unsupervised and supervised machine learning as well as information visualisation. Over the years, such methods have continually progressed towards the ability to capture and analyse the structure and latent characteristics of larger and more complex data. In this thesis, we examine the problem of developing efficient and reliable embedding methods for revealing, understanding, and exploiting the different aspects of the relational data. We split our work into three pieces, where each deals with a different relational data structure. In the first part, we are handling with the weighted bipartite relational structure. Based on the relational measurements between two groups of heterogeneous objects, our goal is to generate low dimensional representations of these two different types of objects in a unified common space. We propose a novel method that models the embedding of each object type symmetrically to the other type, subject to flexible scale constraints and weighting parameters. The embedding generation relies on an efficient optimisation despatched using matrix decomposition. And we have also proposed a simple way of measuring the conformity between the original object relations and the ones re-estimated from the embeddings, in order to achieve model selection by identifying the optimal model parameters with a simple search procedure. We show that our proposed method achieves consistently better or on-par results on multiple synthetic datasets and real world ones from the text mining domain when compared with existing embedding generation approaches. In the second part of this thesis, we focus on the multi-relational data, where objects are interlinked by various relation types. Embedding approaches are very popular in this field, they typically encode objects and relation types with hidden representations and use the operations between them to compute the positive scalars corresponding to the linkages' likelihood score. In this work, we aim at further improving the existing embedding techniques by taking into account the multiple facets of the different patterns and behaviours of each relation type. To the best of our knowledge, this is the first latent representation model which considers relational representations to be dependent on the objects they relate in this field. The multi-modality of the relation type over different objects is effectively formulated as a projection matrix over the space spanned by the object vectors. Two large benchmark knowledge bases are used to evaluate the performance with respect to the link prediction task. And a new test data partition scheme is proposed to offer a better understanding of the behaviour of a link prediction model. In the last part of this thesis, a much more complex relational structure is considered. In particular, we aim at developing novel embedding methods for jointly modelling the linkage structure and objects' attributes. Traditionally, link prediction task is carried out on either the linkage structure or the objects' attributes, which does not aware of their semantic connections and is insufficient for handling the complex link prediction task. Thus, our goal in this work is to build a reliable model that can fuse both sources of information to improve the link prediction problem. The key idea of our approach is to encode both the linkage validities and the nodes neighbourhood information into embedding-based conditional probabilities. Another important aspect of our proposed algorithm is that we utilise a margin-based contrastive training process for encoding the linkage structure, which relies on a more appropriate assumption and dramatically reduces the number of training links. In the experiments, our proposed method indeed improves the link prediction performance on three citation/hyperlink datasets, when compared with those methods relying on only the nodes' attributes or the linkage structure, and it also achieves much better performances compared with the state-of-arts.