TagNet: Toward Tag-Based Sentiment Analysis of Large Social Media Data

Abstract

Hashtags and replies, originally introduced on Twitter, have become the most popular ways to tag short messages in social networks. While the primary uses of these human-labeled metadata are still for message retrieval and clustering, there have been increasing attempts to use them as subject or topic indicators in measuring people's continuous sentiments in large message corpora. However, conducting the analysis for large social media data is still challenging due to the message volume, heterogeneity, and temporal dependence. In this paper, we present TagNet, a novel visualization approach tailored to the tag-based sentiment analysis. TagNet combines traditional tag clouds with an improved node-link diagram to represent the time-varying heterogeneous information with reduced visual clutter. A force model is leveraged to generate layout aesthetics from which the temporal patterns of tags can be easily compared across different subsets of data. It is enhanced by visual encodings for quickly estimating the time-varying sentiment. Interaction tools are provided to improve the scalability for exploring large corpora. An example Twitter corpus illustrates the applicability and usefulness of TagNet.