A bi-directional flow weighted regression for interpreting social media sentiment identified by large language models
Anqi Lin, Hengyuan Liu, Hao Wu
Abstract
Social networks, combined with location-based services, offer valuable opportunities to examine social media sentiment and interactions across regions. Information flow within social networks is often highly directional and intense, transcending geographic distances. As a result, Geographically Weighted Regression (GWR), a traditional model that uses geographic distance to measure spatial proximity, falls short in explaining the factors influencing social media sentiment. To address this limitation, this study proposed a bi-directional flow weighted regression (BDFWR) model, supported by large language models, to interpret influencing factors of social media sentiment. The results demonstrated that the BDFWR model outperformed the GWR model by effectively capturing the relationship between social media sentiment and socioeconomic factors. This approach revealed deeper insights into the spatial heterogeneity of social media sentiment across diverse regions, enhancing the accuracy of modelling social media sentiment distribution. Incorporating bi-directional flow distance significantly improved the model’s performance, particularly in cases involving ‘closely low interflow’ and ‘remotely high interflow’ phenomena—critical aspects often neglected in conventional geographic models. Moreover, large language models excelled in detecting implicit positive and negative trends within textual data, offering a promising avenue for advancing sentiment analysis research.