Sentiment mapping: point pattern analysis of sentiment classified Twitter data
Ken Camacho, Raechel A. Portelli, Ashton Shortridge, Bruno Takahashi
Abstract
Detecting and monitoring collective public opinion via social media platforms can provide real-time information to researchers and policymakers. Human emotions, culture, and opinions can be tracked over time to understand where different sentiments manifest themselves geographically. Expanding on existing methodology, the present study draws from sentiment analysis and point pattern analysis to categorize and analyze sentiment toward natural gas across the United States as a means of applying these techniques together. Three methods of machine learning were used to classify collected tweets into positive and negative categories: Naïve Bayes, Support Vector Machine, and Logistic Regression. Spatial clustering methods and spatial scan statistics were then applied to geocoded tweets to examine the distribution of sentiment about natural gas. In this analysis, the Logistic Regression and Support Vector Machine methods outperformed Naïve Bayes in classifying sentiment. The different methods produced not rather different classification results but also produced varying geographic results. The spatial analyses successfully indicated persistent patterns of negative and positive tweeting about natural gas that correlate with expectations given the physical and cultural environment of various regions. Further, the temporal variation of geographic hotspots of sentiment was readily apparent, suggesting that these approaches can reveal dynamic sentiment landscapes.