Visual analytics for exploring topic long-term evolution and detecting weak signals in company targeted tweets. (October 2017)

• Record Type: Journal Article
• Authors: Pépin, Lambert; Kuntz, Pascale; Blanchard, Julien; Guillet, Fabrice; Suignard, Philippe
• Is Part Of: Computers & industrial engineering. Volume 112(2017)
• Page Start: 450
• Page End: 458
• Publication Date: 2017-10
• Subjects: Social media -- Visual analytics -- Topic mining -- Weak signals -- Twitter; Engineering -- Data processing -- Periodicals; Industrial engineering -- Periodicals; 620.00285
• DOI: 10.1016/j.cie.2017.01.025
• Languages: English
• ISSNs: 0360-8352

Abstract

Highlights: We propose a visual analytics approach to track topics relative to a company from Twitter. The approach combines topic modeling and topic temporal evolution visualization. We perform an experimental analysis of dissimilarity measures to assess topic proximities. The approach has been used by the EDF company to detect previously unknown patterns in Twitter. Abstract: Business decision support tools, including social media data analysis, are required to help managers better understand trends and customer opinions. This paper presents a visual analytics-based approach to assist an expert user in tracking topics relative to his/her company from Twitter. Developed for visualizing topic long-term evolution and detecting weak signals, our process is composed of three complementary steps: (i) a time-dependent topic extraction based on a Latent Dirichlet Allocation, (ii) a topic relationship detection based on a dissimilarity which evaluates the topic proximities between consecutive time slots, and (iii) a topic evolution visualization inspired by a Sankey diagram popular in industrial environments to show dynamic relationships in a system. To test our approach, we have used a real-life dataset from the French energy company EDF from which we have analyzed the evolution of a corpus of more than 70 000 tweets related to this company published over one year, and detected different types of evolving patterns hidden by the data volume and commonly masked by fully automatic mining algorithms.