Sentiment-aware sampling for large-scale geo-tagged social media data abstraction

With the increasing size of social media datasets, it has become a difficult task to explore significant features such as sentiments from traditional geographical visualizations, because the visual elements of data items always overlap heavily with each other. A large number of sampling methods have been proposed to reduce the visual clutter and simplify the geographical visualizations. However, the sentiment features are ignored in the course of sampling which make it easily to misunderstand the sentiments across local areas. In this paper, we propose a visual abstraction framework for the simplification and exploration of large-scale geo-tagged social media data, aiming to preserve both the sentiment features and spatial distribution with the sampled social media data items. Firstly, an agglomerative hierarchical clustering method is proposed to classify social media data into hierarchical categories based on geographical locations and sentiment attributes. Then, a multi-scale rapid sampling model is designed to select data items from categories to guarantee sentiment distribution of the local and global areas in the hierarchy. A rich set of visual interfaces are designed including sentiment tree, heatmap, and ranking view, enabling users to explore and evaluate the simplified geographical visualizations. We also provide several user-friendly interactions such as the sentiment wheel for users to conduct deeply geospatial sentiment analysis of large-scale geo-tagged social media data. Case studies based on real-world datasets and interviews with the domain experts have demonstrated the effectiveness and practicability of our system in simplifying and exploring large-scale geo-tagged social media data.