Activity Discovery Using Large-Scale Geo-Tagged Social Media

Understanding human activities is gradually recognized as a cornerstone task that is critical to a large number of practical applications, ranging from urban planning to location recommendation. While traditional approaches of analyzing people’s city-wide activities heavily rely on paid surveys and human investigations , the explosively growing geo-tagged social media (GSM) brings new opportunities of solving such problems. A key challenge in activity discovery is how to model human activity from the highly noisy and complex GSM data. In this paper, we propose GSM2VEC, an efficient and scalable method that analyzes people’s city-wide activities using large-scale geotagged social media (GSM) data. GSM2VEC consists of three key modules. The first module is a classifier that distinguishes activityrelated records from babbling records. Specifically, we build a cotraining framework that leverages both keyword semantic similarity and spatiotemporal distribution similarity. The second module is a scalable and effective clustering framework that finds a number of micro-hotspots in both the geographical and temporal spaces. The third module of GSM2VEC embeds temporal, spatial and textual units from those extracted micro-hotspots into a unified space. As a result, given any two of the three units, the left unit can be inferred by maximizing its correlation with those two input unit. Our extensive experiments on millions of geo-tagged tweets generated in New York and Los Angeles show that GSM2VEC signifcantly outperforms state-of-the-art methods in all three problems: text prediction, location prediction, and time prediction.