Advances in Learning and Understanding with Graphs through Machine Learning

Abstract

Graphs have increasingly become a crucial way of representing large, complex and disparate datasets from a range of domains, including many scientific disciplines. Graphs are particularly useful at capturing complex relationships or interdependencies within or even between datasets, and enable unique insights which are not possible with other data formats. Over recent years, significant improvements in the ability of machine learning approaches to automatically learn from and identify patterns in datasets have been made.

However due to the unique nature of graphs, and the data they are used to represent, employing machine learning with graphs has thus far proved challenging. A review of relevant literature has revealed that key challenges include issues arising with macro-scale graph learning, interpretability of machine learned representations and a failure to incorporate the temporal dimension present in many datasets. Thus, the work and contributions presented in this thesis primarily investigate how modern machine learning techniques can be adapted to tackle key graph mining tasks, with a particular focus on optimal macro-level representation, interpretability and incorporating temporal dynamics into the learning process. The majority of methods employed are novel approaches centered around attempting to use artificial neural networks in order to learn from graph datasets.

Firstly, by devising a novel graph fingerprint technique, it is demonstrated that this can successfully be applied to two different tasks whilst out-performing established baselines, namely graph comparison and classification. Secondly, it is shown that a mapping can be found between certain topological features and graph embeddings. This, for perhaps the the first time, suggests that it is possible that machines are learning something analogous to human knowledge acquisition, thus bringing interpretability to the graph embedding process. Thirdly, in exploring two new models for incorporating temporal information into the graph learning process, it is found that including such information is crucial to predictive performance in certain key tasks, such as link prediction, where state-of-the-art baselines are out-performed.

The overall contribution of this work is to provide greater insight into and explanation of the ways in which machine learning with respect to graphs is emerging as a crucial set of techniques for understanding complex datasets. This is important as these techniques can potentially be applied to a broad range of scientific disciplines. The thesis concludes with an assessment of limitations and recommendations for future research.