Research On Application Of Deep Learning In Recommendation System

With the rapid development of information networks,numerous and redundant data will always be generated in recent years.Faced with massive amounts of data,users may not be able to accurately select effective information.Therefore,recommendation systems have emerged as a necessary tool to help users obtain accurate and effective information.Traditional recommendation methods are prone to problems such as data sparsity and cold start,and it is difficult to deal with the problems of user preferences and the relationship between users and items changing over time.Matrix factorization and tensor factorization have been widely used in recommender systems,but they still have many limitations.Firstly,small random perturbations of linear model parameters may cause large backward errors and lack of learning the non-linearities between different latent factors.Secondly,it does not consider that the resultant model is highly vulnerable to adversarial perturbations on its model parameters,which implies the possibly large error in generalization.Thirdly,in the absence of key contextual information to capture documents,interaction with useless features may introduce noise and reduce performance.In response to the above problems,this paper combines deep learning with traditional frameworks to build a new deep network for representation learning,which effectively alleviates data sparsity.Adding adversarial perturbations on model parameters stabilizes the model fitting process and improves the robustness of a recommender model and generalization performance.The main work of this paper is as follows:1.For collaborative filtering algorithms combined with deep learning do not consider the problem of multi-dimensional interaction of linked data changing dynamically over time,and propose a tensor factorization recommendation model that combines time interaction learning and attention long short-term memory networks(LSTM-Attention Neural Tensor Factorization,LA-NTF).By using the long short-term memory network based on attention mechanism to extract the latent vector of the item from the item text information,and characterizing multidimensional interactions of user-item relational data in time using the long short-term memory networks based on attention mechanism.Finally,the user-item-time 3D tensor is embedded in the multi-layer perceptron to learn the non-linear structural features between different latent factors,to predict the user's rating of the item.Extensive experiments on Movie Lens-1M and Netflix datasets show that RMSE and MAE indicators significantly outperform neural network based on factorization models and other traditional methods,indicating that the significant improvement in rating prediction task on various dynamic relational data by our LA-NTF model.2.Aiming at the problem that traditional recommendation algorithms using shallow models cannot learn the deep features of users and items,and that recommendation models are extremely vulnerable to adversarial perturbations on its model parameters.This paper proposes a matrix factorization recommendation model that combines time adversarial learning,attention gated recurrent gate networks and attention convolutional neural networks(GRU-AttentionCNN-Attention-Adversarial Matrix Factorization,GA-CA-AMF).Firstly,the gated recurrent gate networks based on attention mechanism is used to extract the user's potential feature vector from the user text information.Then,the convolutional neural network based on attention mechanism is used to extract the item's potential feature vector from the item text information.Finally,adversarial perturbations are introduced on the potential factors of users and items to quantify the loss of the matrix model under parameter perturbations,thereby predicting the user's rating of the item.Experiments on Movie Lens-1M and Movie Lens-10 M datasets show that the proposed model in this paper enhances the robustness of the recommendation model and thus improve its generalization performance,and alleviating data sparsity at the same time.