Research Of Information Diffusion On Social Network Based On User Features

With the development of social network and information technology, online social network (SNS) has become the main source of information and major channel of information diffusion. Recently, the study of information diffusion has been focused on the better understanding of the pattern of user behavior and better management of public opinion. Also, user interests and user profile can be generated for personal social recommendation by the reseach of information diffusion.Nowadays, main efforts of studying information diffusion focus on the following methods:Analytical research, which focuses on the of properties of information diffusion, while might lack of the exploration of the nature of information diffusion. Explanatory research, which tries to reveal the unknown mechanism of information diffusion, while has constrained generalizability due to the subjectivity of modeling process. Predictive research, which puts the effort on predicting the retweet behavior of a user on a specific tweet, while has less support of the incremental property of evolving social network because of the assumption of the static network architecture. Therefore, how to process interactive social data comprehensively, dynamically and incrementally has become an imperative issue of the information diffusion research.Machine learning technology has broken through in many fields concerning with prediction. Learning to rank, LTR in short, is one of the most important model in information retrieval and has gained increasing attention in the field of social network analysis. LTR has achieved acceptable precision when applied to the task of predicting information diffusion, while suffering from some difficulties. Firstly, features selected might not be capable of revealing the of nature of information diffusion due to the unknown diffusion mechanism and the limited aspect of feature generation. Secondly, the dynamicity of evolving social network makes it more difficult to track the evolution of user follower and followee(fans) network and thus requires incremental and multi-dimensional features. Thirdly, the method of generating candidate sets in LTR might suffer from low efficiency when it's applied to the field of information diffusion due to the huge amount of social network users and tweets. Also, the evolving data stream might bring up the challenge of concept drift, which could contaminate the performance of LTR learning algorithm.Therefore, this thesis mainly focuses on the research of information diffusion predicting algorithm based on LTR architecture as following:1. Combining representative features of analytical, explanatory and predictive methods, this thesis extends diffusion-related features into user-based features, relation-based and tweet-event-based features. Experiment results on Weibo and Twitter datasets indicate that enriched diffusion features are capable of improving the prediction precision of information diffusion.2. A time-window based Incremental Mix Feature Generation(I-MFG) is proposed in this thesis. In I-MFG, the static network architecture is substituted by a more flexible dynamic diffusion architecture, which is generated by the flowing-in and flowing-out of information rather than the following-in and following-out relation. Diffusion-related features extention together with the method of incremental learning are utilized to generate time-window based incremental features. Experiment results on Weibo and Twitter datasets indicate that incremental features generated are more representative and flexible, and can improve the precision of information diffusion prediction.3. A time-window based Incremental pointwise Learning To Rank(I-pLTR) is proposed in this thesis, for the demand of the incrementality of social network data and learning algorithm. The method of candidate sets generation in LTR is improved by approximating social network architecture using user behavior architecture, which is more reasonable and flexible for large social network. I-pLTR can incrementally update the learner of current time-window based on the partial instance memory and partial concept memory, which are conserved from the datasets trained ever according to the specified incremental policy. Experiment results on Weibo and Twitter datasets indicate that I-pLTR can increase the efficiency of generating candidate sets, especially when the social network is large and dense. Also, experiment results indicate that I-pLTR have better prediction performance and generalization probability considering other mainstream information diffusion methods.