Research On Recommendation Algorithm Based On Noise Reduction Self-Encoder

With the development of the Internet,the volume and update rate of online information is increasing,and the information overload problem is becoming more and more serious.It is very difficult for network users to obtain the information they need.The recommendation system is a strategy to solve the information overload problem.Has been widely used in various fields.Most of the traditional recommendation algorithms only use the user's rating information to learn the implicit features,which represent the user's preferences and make recommendations.Although the results are good in the experiment,the cold start and the data sparseness often occur in practical applications,which leads to the recommendation results being inaccurate.With the development of artificial intelligence technology,the recommendation system has a new research direction in this field.The recommendation algorithm based on deep learning is also one of the hot research topics.In view of the inaccuracy of the recommendation results in the traditional recommendation algorithm,this paper combines the advantages of deep learning in feature extraction,mainly doing the following research work:1.In-depth research and introduction to the field of recommendation systems,including the research background and current status of the recommendation system,an overview of the classic recommendation algorithm and analysis of its shortcomings,the deep learning model used in this paper and related recommendation algorithms based on this model.2.Aiming at the problem that the recommendation result in the traditional recommendation algorithm is not accurate enough,a recommendation algorithm UTDAE based on tag information and noise reduction self-encoder is proposed.U-TDAE utilizes the advantages of noise reduction self-encoder in feature extraction to extract the deep features of users.At the same time,the activation function of the noise reduction self-encoder is improved to incorporate the label information in the training process.When the data is sparse,the label vector can be trained as the initial input,which alleviates the data sparse problem.The loss function of the noise reduction selfencoder is improved,so that the prediction error has a larger proportion.Experiments on the Movielens 1M dataset prove that the scoring prediction model can effectively improve the accuracy of scoring prediction,and the algorithm improves the recommendation performance to some extent.3.Aiming at the problem that U-TDAE algorithm does not effectively use project related information,a recommendation algorithm WIU-TDAE based on project similarity weight and noise reduction self-encoder is proposed.The item feature vector of the item tag information is extracted by the noise reduction self-encoder,and the item similarity is calculated.The relationship between the projects and the user's interaction with the project will also affect the characteristics of the user's similarity,calculate the project similarity weight and calculate the user similarity according to the weight.The WIU-TDAE algorithm adjusts the proportion of the original user similarity calculation method and the user similarity calculation method based on the project similarity weight in the end user similarity calculation by the balance coefficient,and finally performs the score prediction based on the calculated user similarity..Experiments on different datasets show that the application of WIU-TDAE algorithm to scoring prediction can improve the prediction accuracy,and the algorithm has better recommendation effect.