Feature-rich Networks: When Topology meets Semantics

A network can be enriched with attributes that embed extra information into the nodes. A network can even be enriched with information that encodes different layers of links or that tracks a topological evolution as time goes by. A recent unifying term, that of feature-rich networks, aims to keep all these aspects together within a common denomination and towards a common framework of analysis. The scope of this thesis is three-fold: i) acknowledge all those models that integrate non-structural information into a complex network topology; ii) define new methods (algorithms and measures) for feature-rich network mining; iii) test such methods on applied case studies among different domains. We overview the most influential feature-rich representations for complex networks: Node-attributed, Multi-layer, and Dynamic models. All of them open many challenges for the improvement of classic complex network tasks, like community detection, synthetic network generation, and measures for capturing networked patterns and behaviors. We question these tasks, and we develop new methods for feature-rich networks. In particular, we propose EVA, a node-attributed community detection algorithm; X-Mark, a node-attributed network generator with planted communities; Conformity, for estimating multi-scale mixing patterns; and Delta-Conformity, an extension of the previous one on dynamic environments. Then, we test the proposed methods on different domain-specific applications. In particular, we focus on feature-rich models of cognition and higher-order dynamic social data with semantic annotations on users. Throughout the work, our main focus is to demonstrate that mining augmented network topologies can provide novel insights in many domains, and that methods for feature-rich networks can unearth patterns that are invisible to structural-only and semantic-only data mining.