Machine learning on complex networks : dynamical fingerprints, embeddings and feature engineering

Complex networks emerge as a natural framework to describe real-life phe- nomena involving a group of entities and their interactions, i.e., a social net- work. Furthermore, other problems involve a collection of networks, such as multi-layer, temporal, or brain networks (connectomes). With the increas- ing availability of graph-shaped data and associated meta-data, i.e., node or edge attributes, machine learning (ML) techniques on networks have risen in popularity. However, there are still several challenging issues to be addressed concerning feature engineering, graph representation, visualization, and general graph mining applications. This thesis contributes to bridging the gap between ML and complex net- works. We investigate dynamics on networks such as discrete and continuous- time random walks, in two directions. First, we develop a multi-scale anomaly detection algorithm on attributed networks. We exploit the link between graph signal processing and the Markov stability framework used in community de- tection, for spotting anomalous nodes in multiple contexts of the network. Second, we introduce a generalization of assortativity on networks operating on scalar and categorical node attributes. These coefficients turn out to be useful descriptors to build network fingerprints, so that we can perform super- vised graph classification, i.e., predicting the toxicity of molecules or classifying Reddit discussion threads. Besides, we develop an unsupervised approach to learn graph embeddings from a collection of networks. By learning a non-linear mapping from input graphs to a lower-dimensional feature space, we obtain useful graphs representa- tions used in graph visualization, clustering, and classification, e.g., predicting people’s gender based on their structural connectome. Finally, we propose a principled approach to identify stable biomarkers for schizophrenia diagnosis in the human connectome. From an ML perspective, the problem is stated as an embedded f
(SC - Sciences) -- UCL, 2019