Research On Deep Collaborative Recommendation Based On Marginalized Denoising Auto-enconder

Collaborative filtering has been widely used in recommending systems to solve real problems.In the traditional collaborative filtering recommendation system,learning the potential factor's effectiveness plays the most important role.The traditional collaborative filtering method mainly uses the matrix factorization technique to learn potential factors from the user's item scoring matrix,but the method faces the problem of cold start and severe sparsity.To reduce the impact of sparseness problems,many improved collaborative filtering algorithms make use of the text information of the project,but there is also a sparse problem of text information.Because deep learning model can be a very good feature extraction method in many applications.And many researchers extract the dense features of textual information through deep learning to optimize the collaborative filtering method.However,due to the deep training needs of their own learning,there is often a long time-consuming problem.In order to solve this problem,this paper proposes a deep collaborative recommendation method based on stacked marginalized auto-encoder.First of all,we use the bag vector of the article's text information as the input of the stacked marginalized Auto-encoder.Then,by reconstructing the bag vector after adding noise and calculating the cross-entropy error,the latent vector of the item is obtained while learning the encoding and decoding matrix.Then,the hidden layer feature vector is used as the basic value of the potential feature vector of the project in the probability matrix decomposition target.The user-item score matrix information is used to obtain the potential eigenvectors of the user and the offset values of the potential eigenvectors of the item from the basic values using the matrix factorization technique.Then,using the latent feature vector of the project obtained by the matrix decomposition technique,the automatic encoder is reversely fine-tuned.After the iterative training of both ends the process of the prediction model,the latent potential eigenvectors of the user and the potential eigenvectors of the object are used to carry out the work of the prediction model.In this paper,under the Movielens and Netflix datasets,the experiments are compared and the experimental results are analyzed.The experimental results show that the proposed deep collaborative recommendation method based on stacked marginalized auto-encoder.can accelerate the learning speed of the model effectively,and also improves the recall rate of the model compared with the previous recommendation method.