A Practical and Critical Look at the Problem of Community Discovery in Multi-layer Networks

A core task in social network analysis is to recover communities; that is, to meaningfully partition the network into groups of nodes such that nodes within a group share common behaviors, traits or properties. Driven by the motivation that different types of relationships/interactions among social actors do not exist in isolation but they might depend on each other such that one can lead to the other one, advances on social network analysis suggested that providing meaningful insights about networks should consider these inter-dependencies to avoid misleading or incomplete findings. This leads to the emergence of multi-layer network analysis; that is, an approach where a single network
models different types of interactions among a set of actors. This generalization proposed by the multi-layer framework for modelling social networks has introduced new challenges to the community detection problem. There has been a general tendency by researchers to cope with this complexity by extending some of the already existentmethods used with mono-relational networks. While the logic behind these methods seems to be sound, I claim that the patterns they identify in multi-layer networks are not thoughtfully investigated yet, which makes their definition of multi-layer communities quite vague and hence providing qualitative interpretation for the outputs becomes not an easy task, especially with real-world data. In my thesis, I investigate community detection from a practical perspective while being equally distant from the different
disciplines that researched this topic. I adopt a critical stance towards community detection in multi-layer networks in some cases when the methods used seem to inherit their success and popularity from their original mono-relational implementations rather than their ad-hoc competency in recovering complex patterns in multi-layer networks. I propose different models for multi-layer communities and challenge popular community detection methods to recover these models. I conclude by pointing out to limitations in
these methods and the available layer-coupling mechanisms to recover the ground truth communities in multi-layer networks.