A Cold Start Recommendation Algorithm Based On User Preferences And Trust

In the era of big data outbreak,collaborative filtering recommendation technology has been widely used by many systems because of its ability to efficiently and simply solve information overload problems.However,there are some shortcomings to deal with user cold start,sparse data,and so on.Because the social trust data can effectively compensate for the lack of cold start user rating data,integrating the social trust data into the traditional collaborative filtering recommendation technology is an effective means to implement accurate recommendation for cold start users.The personalized recommendation technology for cold start users,including the multi-dimensional user trust measurement model and adaptive allocation strategy of weight based on comprehensive similarity are mainly analyzed in this thesis.The research results in this thesis are followings:1.For the singleness of trust metrics,a multi-dimensional trust metric model for users by analyzing the principle of social trust mechanism is proposed in the thesis.In this model,the user trust is composed of direct trust and indirect trust.When the being trusted probability of user is applied to collaborative filtering,it not only considers the asymmetry of trust transfer in the local trust network,but also considers the prestige value of the user in the global trust network,so the accuracy of user trust metrics is improved.The model is compared to other typical models on the Epinions data set,the results show that the model can effectively improve the overall recommendation effect for cold start users.2.In order to improve the accuracy of weight distribution in users' synthetic similarity,an adaptive allocation strategy of weight based on comprehensive similarity is presented in this thesis,which selects the most important evaluation indicators for cold start users to optimize: the mean absolute error,and uses the artificial bee colony algorithm to determine the approximate optimal solution of the weight coefficients in the comprehensive similarity.The strategy accords with the user's individual requirements when dynamically assigning scores and trust weights in the composite similarity,thus it can reduce the recommendation error for each cold start user,and the self-adaptability is better.3.Using a Web page method to display personalized recommendations for cold start users.The Django framework is used to design and implement a prototype system for the above models and strategies in the thesis.