Research On Influence Strength Computing And User Behavior Prediction In Social Network

People’s thoughts, actions and feelings are always influenced by others, which we call social influence. Social influence widely exists in the common life, for example, people always decide to watch a movie influenced by their friends and employee may change their working methods influenced by their colleagues and students may decide to purchase a study book influenced by their classmates. In social psychology community, lots of studies have been devoted to understand why and how the individual’s thoughts, actions, and feelings are influenced by others. In recent years, with the explosive growth of online social networks, much attention has been paid to the research on social influence and influence-driven information diffusion as well as the user behavior analysis in social networks from researchers in computer science, social science, and so on.Because in real life, the influence strength between users are different, so how to accurately measure the influence becomes an important question. So far, numerous attempts have been focus on this question and many methods have been proposed, but most of them only consider user’s positive influence and ignore negative influence. For example, in the movie review sites, users will give a score to the movies they have seen and the score may be high or low which decides the influence to their friends are positive or negative. When the user gives a high score to the movie which may promote his friends to watch the movie and we think the influence between the user and his friends are positive. Otherwise, when the user gives a low score to the movie which may decrease the probability of his friends to watch the movie and we think the influence between the user and his friends are negative. So, we should take both positive and negative influence into account.Understanding and predicting the user behavior could be helpful in many applications such as product recommendation, precision marketing and advertisement targeting, but the factors that influence the user behavior are complicated. As previously analyzed, the social influence is a key factor in affecting the user behavior and in addition, user behavior will be influenced by their own interests and preferences as well as the popularity of products. However most of the previous works predict user behavior just from the angle of user preference or social influence and they don’t take all these factors into consideration. So it is meaningful to accurately predict the user behavior by utilizing all factors comprehensively.Although massive works have been done in computing the influence probabilities between users and predicting the user behavior, this paper dedicates to study the problem of existing research which don’t take the negative factors into account in influence probability computing and also not take the comprehensive factors in to account in predicting user behavior, which decrease the accuracy of influence measure and user behavior prediction. The main contributions are summarized as follows:1. A novel Multipolar Factors aware Independent Cascade model (MFIC) is proposed to model the influence propagation in social network. In MFIC, we analyze the behaviors of users by considering both the neighbors’positive and negative influence on them. We design a method based on the EM algorithm to learn the parameters in our model. Experiments are conducted on two real data sets:Flixster and Digg. Experimental results show that the learnt influence probabilities based on our MFIC model can depict the information diffusion more accurately.2. A comprehensive user behavior prediction model IPLR is proposed, which considers all major factors impacting the user behavior:social influence, user preference and product popularity. A learning method in IPLR based on logistic regression is proposed to learn the weights of different factors and predict the user behavior more accurately. Experiments are conducted on datasets to evaluate our method. The experimental results demonstrate the effectiveness of our proposed model and provide interesting insights into the important factors that drive the user behavior.