Empirical Study Of Sentiment-Aware Modeling And Analyses In Social Networking

Web2.0has influenced and changed people’s daily life. Various social networking services (SNS) empower online users to interact with each other in a more efficient and convenient way. In the meantime, the big data generated by SNS users provides challenges and opportunities for research. On one hand, large-scale real-world data that could hardly be obtained in previous; study is available via SNS now. On the other hand, how to mine the large-scale data and apply them to practical applications is a new problem for researchers. The subjective sentiment is considered to play an important role during individuals’decision-making process. As social networking services have attracted much attention of researchers, how to use sentiment-aware information to further analyze and predict users’behavior and opinion in social networking, is a novel and meaningful study.In this thesis, we propose sentiment-aware models for user behavior prediction tasks, and provide the results of a series of empirical analyses. The three prediction tasks that we focus on are:user relationship prediction, personalized topic recommendation, and user-topic sentiment prediction. The empirical results show that the sentiment-aware models can help improve the prediction performance in all the three tasks.The main contributions of this thesis are as follows.(1) We collected data over a period of time from Twitter, one of the most popular SNS, to create a real-world dataset for our experiments. This dataset includes not only users’profiles, but also their social relationship information, and the tweets they posted. We label the sentiment-aware information in the text content of tweets with a powerful and effective sentiment analysis tool.(2) We propose a sentiment-aware model for user relationship prediction. In this task, users’sentimental influences in SNS are mainly considered. We first define and calculate sentimental influence for users, and then divide them into three types based on the calculation results. The sentiment-aware models SA-UFP and SA-RFP using sentimental influence characteristics as new features are finally proposed for two different user relationship prediction subtasks. Experimental comparison shows that SA-UFP and SA-RFP improve the prediction accuracy respectively in the two subtasks.(3) We propose a sentiment-aware model for personalized topic recommendation. In this task, topical opinion distribution characteristics are mainly considered. We present the topical opinion distribution characteristics and a series of analyses on the collected dataset. Furthermore, the sentiment-aware model SDA-TR is formulated based on the hypothesis regarding to topical opinion distribution and user interests. By comparing the SDA-TR model with the state-of-the-art methods, we empirically demonstrate that SDA-TR performs better in the task of personalized topic recommendation.(4) We propose a sentiment-aware model for user-topic sentiment prediction. In this task, the effect of homophily among social friends is mainly considered. After hypothesis testing for vaildating the homophily in SNS, we propose the sentiment-aware model SFMF. Experimental comparison shows that SFMF outperforms the state-of-the-art methods in the task of user-topic sentiment prediction.