Suchergebnisse

Social media users make decisions about what content to post and read. As posted content is often visible to others, users are likely to impose self-censorship when deciding what content to post. On the other hand, such a concern may not apply to reading social media content. As a result, the topics of content that a user posted and read can be different and this has major implications to the applications that require personalization. To better determine and profile social media users' topic interests, we conduct a user survey in Twitter. In this survey, participants chose the topics they like to post (posting topics) and the topics they like to read (reading topics). We observe that users' posting topics differ from their reading topics significantly. We find that some topics such as "Religion", "Business" and "Politics" attract much more users to read than to post. With the ground truth data obtained from the survey, we further explore the discovery of users' posting and reading topics separately using features derived from their posted content, received content and social networks.

Discovering user demographic attributes from social media is a problem of considerable interest. The problem setting can be generalized to include three components — users, topics and behaviors. In recent studies on this problem, however, the behavior between users and topics are not effectively incorporated. In our work, we proposed an integrated unsupervised model which takes into consideration all the three components integral to the task. Furthermore, our model incorporates collaborative filtering with probabilistic matrix factorization to solve the data sparsity problem, a computational challenge common to all such tasks. We evaluated our method on a case study of user political affiliation identification, and compared against state-of-the-art baselines. Our model achieved an accuracy of 70.1% for user party detection task.

Predicting users political party in social media has important impacts on many real world applications such as targeted advertising, recommendation and personalization. Several political research studies on it indicate that political parties' ideological beliefs on sociopolitical issues may influence the users political leaning. In our work, we exploit users' ideological stances on controversial issues to predict political party of online users. We propose a collaborative filtering approach to solve the data sparsity problem of users stances on ideological topics and apply clustering method to group the users with the same party. We evaluated several state-of-the-art methods for party prediction task on debate.org dataset. The experiments show that using ideological stances with Probabilistic Matrix Factorization (PMF) technique achieves a high accuracy of 88.9% at 22.9% data sparsity rate and 80.5% at 70% data sparsity rate on users' party prediction task.