Research On Hybrid Dynamic Recommendation Based On Multiple Neural Networks

Recommendation system is a kind of automatic information retrieval technology to solve the problem of information overload in the Internet large data scene.However Recommendation based on Matrix factorization,one traditional recommendation model,obtains user hidden features by mining user history data and recommends items on the basis of these hidden features to be in accordance with the needs of users.In fact,the features of interest in this kind of learning methods are generally static,and tend to ignore dynamic characteristics in the recommendation system.Dynamic recommendation wants to learn the dynamic changes of interest characteristics to take dynamic factors into consideration in recommender systems and realizes the recommended task which can update over time.It has important practical significance for the real-time recommendation of users and is especially suitable for users whose interest changes fast in the mobile Internet environment.The dynamic recommendation model mainly depends on the temporal data,that is,the user's historical behavior data with time stamp.Therefore,it is helpful to understand the dynamic features of users' interest in order to improve the accuracy of personalized recommendation.The recommendation system based on recurrent neural network can be used to find out the user's historical behavior patterns,so as to find out the rules of user interest transfer.In our thesis,the application of two types of neural networks: recurrent neural network and feedforward neural network,mines user long and short term interests to establish hybrid dynamic recommendation system achieving a balance between short and long-term interest of users,to further improve the recommendation accuracy.In this paper,we study further in these domain and propose some innovation and their research results are as follows:(1)In this paper we researches the dynamic problems of the recommendation system and put forward the concept of user behavior term.Recurrent neural network will be introduced into the field of dynamic recommendation system;user behavior history data can be split into segmentation of user behavior term to learn model parameters about the short-term interest related learning neural network structure of the user.(2)In this paper,we propose the further improvement of the recommendation system model based on recurrent neural network by add historical element composed of the behavior of user term data,used to store the historical term behavior,in order to improve traceability of historical information of recurrent neural network for.At the same time,we put forward to increase the Embedding layer and Dropout layer to alleviate the overfitting problem of neural network training in recurrent neural network.(3)The thesis presents a fusion of two kinds of neural network models-recurrent neural network(RNN)and feedforward neural network(FNN)as multiple neural network hybrid dynamic recommendation model(Hybrid Dynamic Recommendation Model Based on Multiple Neural Networks,referred to as MN-HDRM).MN-HDRM recommendation system is combined with the characteristics of two kinds of neural network of recurrent neural network and feedforward neural network,integrated of short-term interest factors of user behavior under short-term condition and long-term interest factors under the users' global environment,to achieve dynamic balance between long-term and short-term preference in recommender systems.At the same time,in the model we choose the Bayesian personalized ranking(BPR)as the objective function for best recommended items list,in order to achieve the best results in Top-N recommendation.Finally,in this paper we compare MN-HDRM model with other popular dynamic recommendation algorithms: TimeSVD++,recommendation method based on HMM and the recommendation method based on RNN,on two real datasets: Last.fm and Tmall in the experiments.The results of experiments show that the proposed MN-HDRM model shows superior performance in the precision,recall,F-measure and Mean reciprocal rank.