Cross-media User Profiling with Joint Textual and Social User Embedding

Proceedings of the 27th International Conference on Computational Linguistics, pages 1410–1420 Santa Fe, New Mexico, USA, August 20-26, 2018. 1410 Cross-media User Profiling with Joint Textual and Social User Embedding Jingjing Wang1,2, Shoushan Li1,2, Mingqi Jiang1,2, Hanqian Wu3, Guodong Zhou1,2,∗ 1NLP Lab, School of Computer Science and Technology, Soochow University, China 2Collaborative Innovation Center of Novel Software Technology and Industrialization 3School of Computer Science and Engineering, Southeast University, China djingwang@gmail.com, {lishoushan, gdzhou}@suda.edu.cn 20165227010@stu.suda.edu.cn, hanqian@seu.edu.cn Abstract In realistic scenarios, a user profiling model (e.g., gender classification or age regression) learned from one social media might perform rather poorly when tested on another social media due to the different data distributions in the two media. In this paper, we address cross-media user profiling by bridging the knowledge between the source and target media with a uniform user embedding learning approach. In our approach, we first construct a cross-media user-word network to cap- ture the relationship among users through the textual information and a modified cross-media user-user network to capture the relationship among users through the social information. Then, we learn user embedding by jointly learning the heterogeneous network composed of above two networks. Finally, we train a classification (or regression) model with the obtained user embed- dings as input to perform user profiling. Empirical studies demonstrate the effectiveness of the proposed approach to two cross-media user profiling tasks, i.e., cross-media gender classification and cross-media age regression. 1 Introduction User profiling is a task which leverages user generated content (UGC) to automatically identify the data about a user, such as gender (Wang et al., 2017; Li et al., 2016; Li et al., 2015) and age (Marquardt et al., 2014) identification. Recently, along with the boom of social media, user profiling has been drawing more and more attention in various social applications, such as personality analysis (Li et al., 2016; Sarawgi et al., 2011; O’Connor et al., 2010), intelligent marking (Preotiuc-Pietro et al., 2015) and online advertising (Zhang et al., 2016; Volkova et al., 2013). In the literature, conventional approaches normally recast user profiling as a supervised learning prob- lem (Marquardt et al., 2014; Zhang et al., 2016; Ciot et al., 2013; Corney et al., 2002) by exploring various user generated textual features and user social connection features. However, the performance largely depends on a large amount of labeled data and the performance normally degrades dramatically when the model is tested on a different social media. Even worse, many real scenarios may involve several social media with some of them lacking sufficient labeled data to train a well-performed model in each social media. For example, Facebook.com is a social media site where many users publicize their age attribute in their homepages, making the collection of labeled data easy, while Linkedin.com is a social media site where age attribute information is not available in users homepages, making the collection of labeled data rather difficult. These scenarios raise a challenging user classification task, which leverages labeled data from a social media to train a model and …

Content truncated to 500 words