Research On Rating Recommendation Algorithm Of Joint Deep Model

In recent years,with the rapid development of technology,the amount of information presents an explosive growth,leading to users' difficulty in finding the information they want,resulting in "information overload".As common method of information filtering,recommendation system is an effective way in solving information overload problem.The core of recommendation system is recommendation algorithm,which determines the performance of the recommendation system.As a recommendation algorithm,collaborative filtering technology has been used to recommendation systems.However,the effectiveness of such algorithm is limited,it just learns the shallow features of users and items,but cannot do well in the deep feature representations,which restricts the performance of recommendation algorithm.In addition,excessive data will bring about the problem of data sparsity,which will also affect the recommendation accuracy.Deep learning technology has made great breakthroughs in the application of image,speech,and so on.The reason is that this method deep learning-based can learn interactive data through deep non-linear network structure,so as to obtain a unified deep feature representation of the data.Therefore,many scholars and experts have proposed some methods of applying deep learning to recommendation.In this way,it can solve the shortcomings of collaborative filtering algorithm and its extension algorithm can be effectively dealt with.so as to achieve better recommendation accuracy.Because of the insufficient ability of mining implicit information and the problem of data sparsity,these algorithms cannot get satisfactory results.Thus,this dissertation has made some research on existing domestic and foreign rating recommendation algorithms based on collaborative filtering,and we mainly researched on rating recommendation algorithms deep learning-based for mining implicit information,solvieng data sparsity,information utilization,improving recommendation accuracy and so on.The main work of this thesis includes:1.The thesis briefly describes the theoretical basics of recommendation system,collaborative filtering and deep learning,investigating the research status and the shortcomings of rating-based recommendation algorithms at home and abroad.Then,the recommendation algorithm collaborative filtering-based,and deep learning-based recommendation algorithm are introduced in detail.Because of insufficient recommendation accuracy of the current rating recommendation algorithms and so on,utilizing with the current deep learning algorithm framework,we proposed deep hybrid model rating recommendation algorithm and dual-learning based on self-attention rating recommendation algorithm respectively.2.The existing Wide and Deep recommendation algorithm has a good effect when it applied to APP recommendation,but facing poor recommendation accuracy,inability of models to share inputs,and insufficient information mining capabilities when it directly used to rating recommendation.In order to address these problems,we firstly utilize user-item rating history records and share input data.After that,it is directly modified into a model that can mine the ability of hiding user information based on Wide model,not just using the user's memory information ability.A Deep Hybrid Model Rating Recommendation Algorithm(DeepHM)based on the Wide and Deep model is proposed.Comparative experiments on the real datasets MovieLens 100 K and Movielens 1M show that the DeepHM algorithm has a great improvement in recommendation accuracy even if the data is sparise.3.The current research on rating-based recommendation algorithms,including DeepHM,all of them focus on improving recommendation accuracy and information mining capability.But they only adopt the user-item interaction information,and do not consider the implicit interactive information such as user-user and item-item.Therefore,there are still shortage in information utilization and the accuracy improvement.Based on this,we can learn user-user,user-item and item-user information through dual learning mechanism.Then,it combines user self-attention mechanism,item self-attention and interactive self-attention mechanism,which can learn well in different item weights of specific user,so we proposed Dual Learning based on Self-Attention Recommendation Algorithm(DLSA).Except for comparing algorithms without text information of on the MovieLens 100 K and MovieLens 1M datasets,we also compared the comparison algorithm with text information.Experiments show that our proposed DLSA algorithm can achieve better recommendation results under both MAE and RMSE measures.