Research On Attention Collaborative Autoencoder Of Recommender Systems

Thanks to the rapid popularization of smart phones,people can access Internet resources more conveniently.It is difficult for people to choose suitable items from a large number of candidate items under these circumstances.The appearance of personalized recommender systems alleviates this phenomenon.On the one hand,it provides users with suitable products,and on the other,it also helps manufacturers find appropriate consumers.Usually,recommender systems will infer the user's specific likes with the user's rating data,or use some user historical information to recommend suitable item's lists for users.With two kinds of feedback data,this paper uses collaborative filtering as the theoretical idea and selects the denoising autoencoder as the basic network framework.This paper proposes two new neural recommendation frameworks to solve the two kinds of recommendation tasks: 1.Rating prediction : This paper choose explicit feedback data and apply the new recommendation model to obtain predicted ratings,which can reflect the preferences of different users? 2.Top-k recommendation: This paper combine implicit data and build a recommendation framework by a new network,and recommend the corresponding item list to different users The main research work of this paper is as follows:1.Aiming at the task of rating prediction: In order to overcome the shortcomings of traditional recommendation methods dealing with interaction matrix by linear fashion,based on the denoising autoencoder,this paper propose a new neural network model.Due to the structural advantages of the neural model,our model has a special ability to capture the non-linear relationship between user and item.At the same time,in order to handle a large amount of sparse data,this paper propose a sparse forward propagation strategy,which also greatly reduces the number of weights that need to be transmitted in the network.Neural networks that are directly modeling usually do not consider some important effects of users or items in the neighborhood.For this reason,this paper propose an attention strategy to help the model pay attention to special users and items in certain local areas by introducing attention units.In addition,this paper also design a new loss function and apply programming-level marking methods to avoid a large number of loop operations during the calculation,thereby optimizing the training process of our model.2.Aiming at the Top-k recommendation task: in the same way that traditional rec-ommendation methods cannot capture information of user and item in a non-linear manner,this paper still propose a new neural network model for implicit data based on the denoising autoencoder.In order to overcome the shortcomings of sparse implicit data information that is not obvious,this paper propose a sparse data enhancement strategy.By enhancing the data that can reflect the user's real natural preferences,the model has more abundant input data.Compared with the single attention strategy which often used in neural network models under implicit data,this paper propose a multi-dimensional attention strategy to capture the specific preferences of different interactive items for users from different angles.In order to avoid overfitting under implicit feedback,this paper design a new loss function based on item popularity,and enable the model to consider the impact of missing data during the training process.3.Comparative experiments under different public datasets: In order to visually and intuitively demonstrate the effectiveness of our models,this paper design a large number of comparative experiments on two groups of different public datasets.For the rating prediction task,this paper use the square root error to evaluate the prediction accuracy of different methods.For Top-k recommendation,this paper use hit rate and normalized loss cumulative gain to verify the accuracy and relevance about the items between recommendation list by different methods.Finally,this paper verify the effectiveness of each module in our two models through a variety of different types of experiments.