Improving Users’ Demographic Prediction via the Videos They Talk about

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1359–1368, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Improving Users’ Demographic Prediction via the Videos They Talk about Yuan Wang, Yang Xiao, Chao Ma, and Zhen Xiao Department of Computer Science, Peking University, Beijing 100871, China {wangyuan, xiaoyang, machao, xiaozhen}@net.pku.edu.cn Abstract In this paper, we improve microblog users’ demographic prediction by fully utilizing their video related behaviors. First, we collect the describing words of currently popular videos, including video names, actor names and video keywords, from video websites. Secondly, we search these describing words in users’ microblogs, and build the direct relationships between users and the appeared words. After that, to make the sparse relationship denser, we propose a Bayesian method to calculate the probability of connections between users and other video describing words. Last- ly, we build two models to predict users’ demographics with the obtained direct and indirect relationships. Based on a large real- world dataset, experiment results show that our method can significantly improve these words’ demographic predictive ability. 1 Introduction Recent studies have indicated that users’ demo- graphics can be predicted from their linguistic characteristics. A typical practice is cutting the text into a bag of words and training a linear classifier. Although this practice can achieve an acceptable result in simple tasks such as predicting gender and age, it loses some important information about the text structure and does not fully use the relationship between words. Nowadays, people spend a lot of time on videos and social media which provide them with access to post views and comments. Weibo is one of the biggest microblogging platforms in China. More than one third of the “Weibo Trends”1 are about videos. Generally, people with different demograph- ic attributes usually have different tastes for videos (Abisheva et al., 2014). For example, in China people who watch English drama tend to be well- educated. Here is a question: if the video related information in users’ weibo messages can be fully used, will the users’ demographic prediction be improved? One challenge is that many users do not directly mention the video names in their weibo messages. Instead, they make comments on the actors or the plots. If a person likes “Big Bang Theory”, he may post “Will the Big Bang Theory last into the next century?” where the sitcom’s name is mentioned directly, or “Sheldon is so cool, I love him!” which talks about an actor of the sitcom. Both posts indicate the user is interested in “Big Bang Theory”. When involving the demographic prediction, however, the traditional “bag of words based” model cannot extract the above information effectively. Some previous works use topic models such as LIWC (Pennebaker et al., 2001) or LDA (Blei et al., 2003) to detect the relations among users’ words. Usually, they suffer from the short length of weibo messages and the number of topics. In addition, the lifespan of most popular video programs is not very long, which …

Content truncated to 500 words