Deep Learning-Based Estimation of SNS Users’ Place of Residence Using Posts and Following Relationships

In this paper, we propose a new method for estimating the place of residence of Social Networking Service (SNS) users living in Japan. In our previous paper, we proposed a model that estimates the users’ place of residence at the prefecture level based on the content of their posts and the weather and earthquake information described in them. While the weather and earthquake information contributed to an improvement in precision, the information did not enhance recall, particularly in prefectures with a small number of users or those adjacent to large metropolitan prefectures. Therefore, we need to improve the recall for these prefectures in this task. To address this issue, we focus on that the users living in rural areas tend to have closer relationships than those in urban areas. We aim to improve recall, especially in rural prefectures, by utilizing the following relationships between SNS users. In addition, we introduce Bidirectional Encoder Representations from Transformers (BERT) model as feature extraction from tweet content to further improve precision. In our evaluation experiments using the same dataset as in our previous study, we achieved an F1-measure of 0.7747. This is the best result in our study so far. Furthermore, we observed an improvement in the recall of up to 5 points compared to the baseline model using only the tweet content.