| With the popularization of mobile internet and smart terminal, network resources presents exponential growth trend. The rich network resources become a double-edged sword, the huge amount of no-needed resources with little needed resources, namely "information overload" problem. The search engines is the mainstream solution of this problem, but it is the indiscriminate service based on keywords that different persons will get the same results using the same keywords using search engine. This makes users experience worse. With user profiles models, different persons can get different results according users profile, and user profiles modeling is the key node of recommender systems. However, there are some drawbacks of the traditional user profiles models. (1) With the popularization of mobile internet, the traditional topics extraction algorithm can not handle short texts as microblogs, comments or short messages accurately; (2) The interests and behaviors of users have effection between users, and the historical behaviors of users have effection on the current behaviors of users. But the user profiles models has not been an in-depth study at present; (3) The traditional user profiles models do not differentiate the difference between long term profiles and short term profiles. However, they have different features, namely that long term profiles are the inherent profiles which are stable and entrenched, conversely, the short term profiles are the profiles appearing because of the effection of neighborhood which are not stable and temporary. If the two kinds of profiles can be considered in different ways, the prediction accuracy and recommend efficiency can be improved greatly.To solve above problems, firstly, this paper proposed an improve LDA microblogs topics model based on weight microblogs chain to extract topics of users microblogs which are short text; secondly, this paper used relationship strength between users stand for the influent between users,and proposed an improved latent factor relationship strength model based on hawkes process; finally, this paper proposed a prediction model of microblog users profile based on users influence to differentiate the long term profiles and short term profiles. The main work and achievements are as follows.1. An improve LDA microblogs topics model based on weight microblogs chain was proposed. To solve the problem that short texts had very low information density, this paper proposed a weight microblogs chain structure that distributing weight according to microblogs published time and social activities information including publish, comment and retweet activities, and took background knowledge to enrich semantic features of this structure, and an improved LDA topics model based on this weigtht microblogs chain structure(WMC-LDA). The experiments showed that this model had smaller Perplexity than standard LDA, namely this model has low predict uncertainty.2. An improved latent factor relationship strength model based on hawkes process was proposed. Aiming at the problem that the traditional researches focus on modeling simple binary relations and static relations, this paper proposes an improved latent factor relationship strength model based on Hawkes process(HP-LRS), which takes the relationship strength, similarity and history interaction behavior between users as a latent factor, latent factor’s incentive and presentation respectively, and uses Hawkes process to characterize relationship between history interaction behavior and users’relationship strength, and this model can be a solution of the disadvantages of the original model without considering the users’history interaction effects and their attenuation. This paper uses the data from microblog social network evaluating HP-LRS model, and the experimental results show that, this model can improve relationship strength prediction accuracy and coverage and sorting accuracy of the Top-N neighbor nodes based on relationship strength.3. A prediction model of microblog users profile based on users influence was proposed, and RBF neural network was used to train the model. For the disadvantages that the existing prediction models without considering the neighbors’ influence and the difference between long term profiles and short term profiles, we propose a prediction model of Microblog users’ profiles, which is based on users influence to improve the prediction precision and enhance the recommendation efficiency of products and service. The model divides the profiles into long term and short term profiles, and users’ influence into themselves and neighbors’. Based on this, we adopt the RBF regularization network which has predominant learning ability and closeness to predict users’ profiles. In the experiment on Tencent Weibo, about 4.31%,14.53% average prediction offset and 0.31,48.12 average variance of prediction offset for long term profiles and short term profiles respectively are achieved, and the result shows that this method is superior to others in prediction precision and stability. |
