Twitter user geolocation based on heterogeneous relationship modeling and representation learning

Abstract

Listen

Translate article

Twitter user geolocation has been garnering considerable attention from academia. Due to the complexity of the Twitter data, the user geolocation performance is limited for some user geolocation methods. Previous works on Twitter user geolocation typically model the user location based on homogeneous relationships, while neglecting the heterogeneous relationships. In this paper, we propose a novel Twitter user geolocation method based on heterogeneous relationship modeling and representation learning. In this method, two heterogeneous graphs are constructed according to the statistical characteristic of the mentioning relationship between users and words, and the mentioning relationship between users in the tweets, respectively. By sampling the nodes in the graph, the complex topological structure of the constructed graph is captured and natural language-like node sequences are generated. The Skip-gram model is used to construct the objective function, and the stochastic gradient descent algorithm is employed to optimize the objective function to learn the representations. Finally, a neural network is used to train the user geolocation model. Experiments conducted on two real-world Twitter datasets demonstrate that the proposed method outperforms the state-of-the-art methods.