A unified approach to semantic and topological similarity in information networks

With the pervasive diffusion of digital social interactions, network analysis became a critical task, as an increasing amount of economic, personal and physical transactions is digitally mediated, recorded and monitored, and can be used to predict and influence the way people or entities interact, work and, in general, behave. A massive quantity of information is generated daily, using heterogeneous multimedia data, by the spontaneous collaborative activity of whom we will just call users in networks. Many systems in nature can be modelled as complex networks i.e., structures consisting of nodes and vertices connected by link or edges. Complex network science is applied to various disciplines, offering a common language to let them interact, e.g. computer science, sociology, biology, co-authorship networks. What distinguishes network science from graph theory is the empirical nature, focused on data. The general idea is to use collaborative/social repositories with the information about entities/users and relationships, as an information source to infer semantic knowledge and relations among heterogeneous multimedia objects of any kind, with the aim of extracting relevant semantic contexts to be then proposed to users. Contexts can also be used to narrow domains and improve performances of tasks. Both web-based and structural similarity measures can profit from suboptimal results of computations, where approximations to evaluate frequencies and probabilities can be used to calculate semantic proximity (i.e. similarity or distance). Which measure to use, and how to optimise the extraction and the utility of the extracted information, are still open issues. The ability to understand the relevant semantic contexts underlying a situation is a significant issue for machine intelligence applications. Semantic similarity measures and topological measures can be tools for evaluating the two classes of similarities. The research questions arise: how to use semantic and topological properties together? How can they be integrated? Is it correct to distinguish between these two classes? Are they really always demanding a different approach, or there are some shared aspects which can be scientifically recognised and formalised? Intuitively, topological similarity may convey also semantic information, e.g. a friendship link may be the emergent signal of an underlying real-world process of interaction between humans. The likelihood that a new friend will be introduced to a person through a common neighbour can be measured as a topological property, but it is motivated by the transitivity of human relationships, which is a real-world process, which can be represented with words in a language. The relation between the semantic and topological aspects in evaluating similarity of network entities, of subnetworks (i.e. communities), networks, and meta-networks is of great interest for our research.