Social User Geolocation Method Based on POI Location Feature Enhancement in Context

The geolocation information in social media user generated text is one of the important bases for inferring users' location. The existing methods often rely on the mentioned Point of Interest (POI) and location Indicative word in user generated text as the inference basis, in consideration of the lack of consideration of semantic information and the inaccuracy or incompleteness of POI contribute to the inaccuracy of user geolocation based on text. To address these issues, this paper proposes Social User Geolocation Method based on POI Location Feature Enhancement in Context (UGLE). Firstly, the POI in user-generated text is removed by deep neural network. Then, through three geographical traits of POI: location orientation, word frequency and word length, the problem of ambiguous POI location orientation in user generated text is solved. Finally, by combining the semantic information of the words around POI, the user vector with enhanced location features is generated, which improves the geolocation accuracy of social users. On the Weibo dataset, the proposed Method is compared with four existing typical methods, PaQL, SeFG, MNB-PART, and MNB-LGR. The results show that the proposed method's accuracy in inferring province-level and city-level locations is 76.78% and 62.56%, respectively. These results are 6.81% and 6.66% higher than the existing best results.