Diachronically like-minded user community detection

Abstract

Diachronically like-minded user community detectionStudy of users’ behaviour, interests, and influence is of interest within the realm of online social networks due to its wide range of applications, such as personalized recommendations and marketing campaigns. However, the proposed approaches are not always scalable to a large number of users and a huge amount of user-generated content. Community-level studies are introduced to facilitate scalability, among other characteristics, highlighting the main properties of the network at a higher collective level. Prior work is mainly focused on the identification of online communities that are formed based on shared links and/or similar content. However, there is little literature on detecting communities that simultaneously share topical and temporal similarities. To extract diachronically like-minded user communities who have similar temporal dispositions according to their topics of interest from social content, we put forward two approaches: i) multivariate time series analysis, and ii) neural embeddings. In the former approach, we model users’ temporal topics of interest through multivariate time series, and inter-user affinities are calculated based on pairwise cross-correlation. While simple and effective, this approach suffers from sparsity in multivariate time series. In the latter method, however, each user is mapped to a dense embedding space and inter-user affinities are calculated based on pairwise cosine similarity. While the objective of these two proposed approaches is to identify user communities up until the present; in the last step of this thesis, we propose two approaches to identify future communities, i.e., community prediction: i) Granger regression, and ii) temporal latent space modeling. In Granger regression, we propose to consider both the temporal evolution of users’ interests as well as inter-user influence through the notion of causal dependency. In the latter method, however, we assume that each user lies in an unobserved latent space, and similar users in the latent space are more likely to be members of the same user community. The model allows each user to adjust her location in the latent space as her topics of interest evolve over time. Empirically, we demonstrate that our proposed approaches, when evaluated on a Twitter dataset, outperform existing methods under two application scenarios, namely news recommendation and user prediction.