Research And Application Of User Physical Attribute Inference Based On Multi-source Network Information Fusion

With the rapid development of social media networks,it is crucial to construct users' portraits from their dynamic behaviors to address the increasing needs for customized information services.Users participate in multiple social media networks to satisfy the diversified demands.The same physical attribute can be inferred from behaviors on multiple social media networks.Fusing multi-source information is conducive to inferring the user attribute to provide better customized information services.However,there are still three problems in multi-source user physical attribute inference: the contradictions in the data,the lack of data labels and the lack of data sources.Therefore,this paper studied and solved the three problems encountered in multi-source user physical attribute inference and evaluated its actual effect in the application.The research contents and major contributions of this paper are as follows:(1)We proposed a multi-source user physical attribute inference algorithm based on hierarchical auto-encoder.It aims to solve the following two problems in multi-source user physical attribute inference: 1)There are contradictions in the data: there is a contradiction between dynamic behaviors and stable physical attributes,and there may be a contradiction among the same individual's behaviors on different social media networks.2)Lack of data labels.The basic idea of this paper is that: the shared pattern among the same individual's behaviors on different social media networks can indicate his/her consistent and stable physical attributes.Hierarchical auto-encoder is introduced in this paper to realize this idea by discovering the non-linear correlation among different social media networks to obtain the shared patterns among the behaviors.The unsupervised scheme in shared pattern learning alleviates the requirements of hierarchical auto-encoder for the overlapped user account and improves the practicability of the model.SVM is then utilized to infer user physical attributes from the derived shared patterns.The experiments on the real-world datasets from three social media networks demonstrated the effectiveness of the proposed method.(2)We proposed the multi-source user physical attribute inference algorithm based on adversarial learning.It aims to solve the problem of the lack of data sources in multisource user physical attribute inference.Based on the discovery that the mapping results of behaviors on different social media networks in the unified user attribute space are independent of the kinds of the social media networks,the proposed approach also simultaneously learns a network source classification network in an adversarial setting to alleviate the requirements of hierarchical auto-encoder for the training data.All users,including single-network users can be used to discover the correlation between the dynamic behavior feature space and the unified user attribute space.It can still take advantage of the nonlinearity of hierarchical auto-encoder in shared pattern learning when the data sources are lack.The experiments on the real-world datasets from three social media networks demonstrated the effectiveness of the proposed method.(3)The application of multi-source credit evaluation.It aims to evaluate the actual effect of multi-source user physical attributes inference.Instantiate the user physical attribute as the Weibo Sunshine Credit,and crawl behaviors on two Chinese social media networks,Weibo and Net Ease Music.Use the two-layer ensemble learning framework to evaluate the user credit.In the first layer of the two-layer ensemble learning framework,it uses hierarchical auto-encoder proposed in this paper to solve the contradictions in the data in an unsupervised way.The accuracy of credit evaluation has been improved.The actual effect of multi-source user physical attribute inference is evaluated by the specific application.