Social Network Users' Personality Traits Mining And Its Application To Personalized Recommendation

The rapid development of social networks provides an excellent opportunity and platform for analyzing user behavior. Topic preference mining based on user-generated content (UGC) attracted research scholars from various fields and disciplines. However,current researches mainly focus on the precise mining of the users' topic preferences and ignore exploring the formation mechism of topic preferences from the perspective of the users' internal factors which is why users have such interests. Personality trait as a typical user intrinsic factor has been widely used to explain the real behavior of human society. Therefore, how to explain social network users' topic preferences from their personality traits is of great theoretical significance and practical value.This thesis focus on social network users' personality mining and its application to personalized recommendation. Specifically, by defining the personality traits as the user's big five personality traits (also called five factor model), this paper examines three progressive scientific problems: analysis of the relationship between personality traits and users' topic preferences based on the nonparametric hierarchical Bayesian topic model; building a recognition model of social network users' big five personality trait based on such relationship; on the basis that the social network users' personality traits can be reasonably recognized, this paper studies the application of users'personality traits in specific circumstances, which is, whether it helps to improve the accuracy of personalized recommendations.The research content and contributions of this thesis include:(1) The analysis of relationship between social network users' personality traits and topic preferences. Based on the assumption that user's topic preference is a result of the user's various personality traits, this paper proposes a novel nonparametric hierarchical Bayesian topic (NHBT), which constructs a three-level generation framework to study the relationship between users' internal personality traits and topic preferences. More specifically, NHBT can accomplish data-driven topic mining task, integrate the topic mining and relationship discovery into a unified model, and allow for the assumptions about "personality traits generating topic" and the multiple-to-multiple relationship between personality traits and users' topic preferences. NHBT model avoid the traditional LIWC dictionary and abandon the traditional two-stage mining task. In this paper, the NHBT model is firstly solved based on the three-level Chinese restaurant process and then a direct sampling method is proposed based on the minimal path assumption. The experimental results on the Facebook dataset show that the NHBT model can explore interesting latent topics from an open social media environment,such as music bank, chemical biology, Cosplay, and so on and can explain the inherent mix mechanisms of users' topic seleation behavior. For example, users of low conscientiousness and high openness prefer to publish topics related to campus life.(2) Personality recognition based on users' topic preferences. Based on the association between social network users' personality traits and topic preferences, a novel probabilistic topic model (PT-LDA) is proposed to solve the problem of personality recognition of unknown users. The PT-LDA extends the Latent Dirichlet Allocation and then reduces the thousands of N-gram features to a number of latent topics. At the same time, it is supposed that each topic not only corresponds to a multinomial distribution over words, but also corresponds to five Gaussian distribution on personality trait value. Since the model is more complex and cannot allow for exact inference and parameter estimation, a Gibbs-EM algorithm is proposed to solve the PT-LDA model iteratively, which is, to execute Gibbs sampling and Expectation maximization alternately. The quantitative evaluation results show that the proposed PT-LDA model is more accurate, efficient and robust than several benchmark algorithms. In addition, even without applying personality traits to "guide" topic mining process, the forecast results of topic feature sets extracted by the standard LDA model are better than that of LIWC feature sets, which further validates the effectiveness of data-dr:iven topic mining method.(3) Integrating personality traits into personalized recommendation. In view of the solid theoretical basis of the close relationship between the personality traits and the users' topic preferences, and the realistic conditions for recognizing the users'personality traits based on such relationship, this paper studies whether personality traits help to improve the accuracy of personalized recommendation and then proposes the personality trait matrix factorization (PTMF) model which fuses the users' personality traits. Specifically, this paper constructs a user-item matrix and a personality trait-item matrix and utilizes the joint decomposition so that personality traits can help to obtain more accurate hidden factor matrices of users and items. In this way, it can alleviate the sparseness of user-item matrix. The experimental results on the music dataset show that the personality traits can effectively solve the problem of data sparseness.This paper expands the research area of the relationship mining between social networks users'personality traits and topic preferences, enriches the modeling methodologies of personality recognition, and provides solid theoretical foundation for applying personality traits information to various practical applications.