Article Preview
Top1. Introduction
Predicting human mobility is a fundamental question for a broad range of applications (Song et al., 2010), including urban planning (Yuan et al., 2012), epidemic forecasting (Dalziel et al., 2013), product advertising (Kirchner et al., 2012) and so on. It is vitally significant to understand socioeconomic phenomena embodying spatiality and human movement by unfolding human mobility patterns (Yan et al., 2013). Therefore, numerous researchers have attempted to uncover and model human mobility patterns (Gonzalez et al., 2008; Noulas et al., 2012; Hasan et al., 2013a; Schneider et al., 2013; Barchiesi et al., 2015; Pappalardo et al., 2015; Gallotti et al., 2016), to provide a deeper understanding of individual and collective mobility behaviors. In recent years, the rapid advances in mobile computing and social networking services empower people to share and use their locations and location-related content in location-based social networks (LBSNs) (Bao et al., 2015) such as Foursquare (https://foursquare.com) and Gowalla1. As a new type of data source containing extensive geo-tagged data with high position resolution and at large spatial scales, these LBSNs have enabled more exciting and critical studies on human mobility than those traditional ways of mobile phone and taxi (Noulas et al., 2012; Barchiesi et al., 2015; Hasan et al., 2013b; Wang & Stefanone, 2013; Wu et al., 2014; Zhang et al., 2015a; Zhang et al., 2015b; Huang & Wong, 2015; Hess et al., 2016).
In statistical physics and computer science, most of the previous studies conducted on large-scale check-in data from LBSNs focused mainly on intra-urban human mobility patterns and dynamics. In addition to the models proposed in statistical physics (Noulas et al., 2012; Hasan et al., 2013a; Schneider et al., 2013; Wu et al., 2014; Huang & Wong, 2015), machine learning based prediction models for points of interest (POIs) have attracted much attention, along with the increasing popularity of recommender systems (Hess et al., 2016). To the best of our knowledge, many types of information are considered as features when training such prediction models (Bao et al., 2015; Hess et al., 2016). These features include user’s check-in history, content information (such as POI properties, user preferences, and sentiment indications) (Gao et al., 2015; Majid et al., 2013), the geographical influence based on the distribution of geographic distance (Ye et al., 2011; Liu et al., 2015), temporal influence (Yuan et al., 2013; Hosseini & Li, 2016), and social ties between friends (Ye et al., 2011; Zhou et al., 2015; Huang et al., 2015). In particular, a few elaborate hybrid models (Zhang et al., 2015b; Gao et al., 2015; Liu et al., 2015; Hosseini & Li, 2016; Huang et al., 2015; Yin et al., 2016), which utilize various combinations of the information mentioned above, have also been made to achieve higher recommendation accuracy for commercial purposes.