TT-graph: A new model for building social network graphs from texts with time series

• Published in: Expert systems with applications Vol. 192; p. 116405
• Main Authors: Jia, Wei, Ma, Ruizhe, Yan, Li, Niu, Weinan, Ma, Zongmin
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 15.04.2022; Elsevier BV
• Subjects: Network analysis; Semantics; Sequences; Similarity; Social network analysis; Social network graph; Social networks; Text semantics; Texts; Time series; User similarity; Text semantics; User similarity; Social network graph; Time series
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.116405

•Proposing a topic model to discover the topic words from time series text.•Exploiting a new method to evaluate the similarity of text with time series.•Combining text semantic and time series to build graph without any prior acknowledge.•Demonstrating the model’s excellent performance in community detection. Social network analysis is a fundamental problem inherent in various applications, which can be handled mainly based on a graph model given in advance. However, it is generally ignored in most existing studies that the social networks may only contain users and no users' connections are explicitly available. Then the connections between users play a considerable role in building the graph, and it is necessary to calculate users' similarities. Traditional methods of user similarity calculation are extensively based on the topics of text content because they can effectively reflect users' interests. Nevertheless, these methods mostly ignore the importance of time series in the texts, where the texts with time series can reveal the activity trends of users in social networks. In this work, we explore a new problem of building a social network graph over text with time series. Our basic idea is that social media users are more similar if they have similar text semantics in similar time sequences. To obtain the semantics of text with time series, we extract topic words of each user from the corresponding text with our proposed Time-Biterm Topic Model (T-BTM), which improves the BTM model by taking the time-topic distribution into account. On this basis, we further propose a novel time series-based graph model with text, called Text with Time series for Graph (TT-Graph) model, which explicitly considers the user similarity and time series similarity. With the TT-Graph model, we propose novel methods for topic detection, community detection, and link prediction in social network analysis. Extensive experiments demonstrate that topic detection, community detection and link prediction can be effectively conducted on the TT-Graph model, and the credibility of our model can be proved.