One-class Collaborative Filtering Recommendation Algorithm Based On Probabilistic Matrix Factorization

The recommendation system,a kind of information filtering technology,can quickly locate the items of interest to the users and solve the contradiction between information overload and user preferences.With the increase of the amount of information,the traditional entity industry has begun to incorporate the recommendation system into their own system to solve the problem of information overload.In the information distribution system of South-to-North Water Diversion,it is very difficult to establish a feature extraction method and feature description model due to the variety of domain information and different data structures.What's more,user interest varies from person to person and can't be well described with the static information.The user implicit feedback behavior is kind of user preference information which is easy to extract and widely exists in various modules of the system.The collaborative filtering recommendation system based on implicit feedback is also called one-class collaborative filtering(OOCF).Therefore,it is of great practical significance to study OOCF recommender system in the information distribution system of South-to-North Water Diversion.However,the OOCF based on matrix flactorization is imbalance due to the lack of negative feedback information,the model need to manually add negative sample to balance.In this paper,a negative sample selection method based on probability sampling is proposed to select the infromation as negative samples that are not browsed.The method treat the non-browsed items of a user as population and think the way to select the negative samples is to sample the population.The characteristics of negative samples include the degree of correlation between users and items and degree of their popularity which can be merged into the process of negative sample selection through the method.The probability sampling method based on weighted feature fusion(PSBFF)is proposed to combine the correlation degree and the popular degree into the same negative sample sampling process.The dataset can be regarded as a kind of two valued data after manually added negative samples and it is not appropriate to model the user behavior with the traditional Probabilistic Matrix Factorization(PMF).A Probabilistic Matrix Factorization based on Binomial Distribution(PMFBD)to solve this problem.Firstly,the Binomial Distribution model is used to replace the Gauss distribution model to establish the interest model of user implicit feedback behavior,and the Logistic regression function is adopted which will limit prediction value between 0 to 1,finally the training method of stochastic gradient decent of the PMFBD model is given.Experiment results show that the probability sampling method with the degree of correlation and degree of popularity has a better performance over the way with random selection and the way based on similarity between items.secondly,the two stage probability sampling method which integrate the two negative samples into the same probability sampling process further improve the overall performance of the recommendation system.Then,by comparing the performance under the different regularization coefficients between the PMFBD model and the PMF model,we can get that the PMFBD model has a better performance than the PMF model when they both have a good regularization coefficients.Finally,we use the PSBFF to.select the negative samples and use PMFBD model to construct user interest model as the method to compare with the traditional method.The experiment shows that the method is better than the traditional method.