Research On Computation Model Of User Influence Based On Feature Learning In Social Network

In recent years, with the rapid development and amounts of applications of internet, online social network fast becomes popular. Social influence of certain users plays a crucial role in various applications of social network such as viral marketing and recommender systems. Therefore, the research on analysis and computation of social influence is urgent required.So far there are lots of research works on social influence, but current methods always implicitly represent influence between users by statistics of social relations, which are lack of deep analysis. Different from these methods, the approach in this paper explicitly model the probability of social influence between users, substantially analyze social influence form the source and propagation. This approach fundamentally attempts to answer why users are influenced and whether it can infer the clues from existing influences in the entire network. Though this method, it will be able to not only describe existing influences but also predict the influences given new users and topics.In this thesis, the main work are summarized as the following three points:1. Give a comprehensive definition for user influence in social network from perspectives of both micro-level and macro-level. Introduce the current two kinds of approaches:descriptive and predictive social influence analysis, and analyze the existing problems.2. Propose a novel influence feature learning approach based on content and social relations, which is able to explicitly model micro-level influence probability and learns a general feature representation for influence. Moreover, deploy an asynchronously paralleled stochastic gradient descent algorithm to learn the predictive influence features by efficiently applying this algorithm on real-world social network.3. Through extensive experiments on a real-world social network with 53 million users and 838 million tweets, the results show the proposed approach significantly improved performance as compared to other state-of-the-art methods in both micro-level and macro-level influence computation and prediction tasks.