Research On Personality Prediction Method Of Social Network Users Based On Multi-label Learning

As the rapid popularization of Internet, social networks have begun to play an important role in public life. People interacting with each other through social networks is becoming an important means of communication. With social networking platform, they express opinions, contact friends and discuss public issues, etc. Social networks have become an extension of real society. There are a large number of frequent activities of users on the public social networking sites(such as Facebook)every moment, either browsing information, or updating the status. As more and more needs of users for social networks, how to provide personalized service has become a hot research topic on intelligent social networking platform, such as friend recommendations and product promotion. As one of the critical factors that affect the user behavior, personality traits can have an important influence on the quality improvement of personalized service. Analysis and prediction of social network user’s personality has board application prospects.In the field of personality psychology, there are many different personality genres. Among them, the personality trait genre provides a possibility to do a relatively reliable and scientific analysis and quantification. The most dependable and popular model is the Big-Five personality model in trait genre. It describes a person’s personality from five aspects. They are Extraversion, Neuroticism, Agreeableness,Conscientiousness and Openness. Big-Five personality model suggests that personality is composed by a variety of characteristics, and its structure is relatively stable. Big-Five personality is close to the behaviors of people in their life, but also has a strong correlation with their network behaviors. With the network mining techniques, through the establishment of computational model between network behavior characteristics and personality traits to achieve the prediction of personality traits via social network information is feasible. In recent years, the studies of socialnetwork user’s personality prediction began to appear. Compared to self-report questionnaire personality calculating means, automated personality prediction using social network information is convenient and objective. Researchers make the feature extraction using web text information with other relevant information available, and then adopt different machine learning techniques, such as k NN, SVM, Naive Bayes and decision tree learning algorithm to build personality prediction model.Experimental results show that the user’s personality can be effectively and automatically predicted based on the social network information. However, the accuracy of prediction of related work is not very good. Better automated predicting methods are needed to fit to the user’s personality prediction, and it’s essential to further tap features that are highly relevant to personality from user-generated information on the social networking platform, and explore internal relations between personality traits. Based on the above problems, the following work has been done:According to user’s text information on social networking sites, we present the design that combining formal features based on words and semantic features for the prediction of personality. Formal features include word feature based on information gain, sentimental, tenses, part of speech and writing style features. Based on correlation with features and the class label collection, MLFSIE-W algorithm is used for feature selection and weighting. Through word mapping based on Word Net ontology, we define the concept vectors, and give an idea of the semantic relevancy computation method combines semantic distance and semantic coincidence degree.According to semantic relevancy and the similarity based on word feature, we present integrated similarity calculating method. In the experiment, various features based on the same classification mathod were compared with related works, and then we discussed the effect of formal features and semantic features in the prediction of personality.Related research on the prediction of user’s personality usually uses the single classification machine learning algorithm for processing. This paper adopts a multi-label user’s personality prediction method based on random walk model for analysis and processing. According to the analysis of Big-Five personality, socialnetwork user’s personality prediction essentially should belong to a multi-label classification problem. In the process of performing a random walk model algorithm,use the integrated similarity to improve its original weight of edge based on Euclidean distance calculating method to construct random walk graphs, through iteration and transformation, then get the probability distribution of each user belonging to each class label, combined with calculating the threshold value, finally make multi-label predictions. Experiments show that the method has better prediction effect than using the single classifier such as SVM, k NN and NB etc. The method takes the potential correlation between each class label into account and the prediction is more reasonable.Aiming at the results of the evaluation of related research on social network user’s personality prediction are generally not ideal, this paper presents using an ensemble learning approach, combining the random walk model, and gives the user’s personality prediction method based on ensemble multi-label learning. In the ensemble approach Ada Boost.MH framework for multi-label learning, change the original idea that resolves multi-label classification into many binary classifications as a single group iterating simultaneously, to directly use the multi-label classifier based on the random walk model as the base classifier. This method not only remains correlation between the classes in the level of the base classifier, but achieves the purpose of ensemble learning. The results show that for user’s personality prediction,ensemble multi-label learning method is effective, and able to further improve the evaluation index of prediction on the basis of multi-label base classifier to achieve better prediction.