Research On Neural Collaborative Filtering Recommendation Algorithm Based On Auto-encoder

In recent years,with the continuous development of Internet technology,the amount and complexity of data of information has increased rapidly.Meanwhile,information overload has become a significant problem in our the society.As an important method to solve information overload,recommendation algorithm has attracted attention of academic and industrial researchers all the time.Most of the traditional recommendation algorithms are designed for explicit feedback.Nevertheless,owing to the defects of data of implicit feedback,the performance of implicit feedbackoriented recommendation algorithm is not good enough.With the development of deep learning,the combination of deep learning and recommendation algorithm has become a hotspot in the research of recommendation algorithm.To be specific,model of neural collaborative filtering is one of the representative works.This model overcomes the shortcomings of data of implicit feedback through neural network,and achieves better performance.However,neural collaborative filtering has the following shortcomings.First,in the data processing stage,the hidden vectors generated by the model carry less effective information and are not capable of representing the essential attributes between users and items adequately.Second,this model uses multi-layer perceptron neural network witch has too many parameters,resulting in a long convergence time.Therefore,it is not suitable for recommendation scenarios with higher timeliness requirements.1)Focusing on the problem that the hidden vectors carry less effective information in the neural collaborative filtering model,this thesis proposes a neural collaborative filtering model based on the variational auto-encoder.Specifically,the hidden vectors of users and items are generated by the variational auto-encoder.Compared with the simple way of one-hot coding,variational auto-encoder generates stable random variable distribution function by utilizing users' historical behavior records,which effectively eliminates data noise and data redundancy in data of implicit feedback.Therefore,the implicit vectors generated by this distribution function are capable of better representing both users and data.In addition,this model redesigns neural network structure as well as integrates the linear and non-linear features between users and items at the bottom of the network,so that it can extract the potential feature between users and items well.Compared with the mainstream recommendation model,the performance of the neural collaborative filtering model based on the variational autoencoder is improved by an average of 2.5%.2)To solve the problem of long convergence time within neural collaborative filtering model,a convolutional neural collaborative filtering model based on denoising Auto-encoder is proposed.This model adopts a simple structure of denoising Auto-encoder to generate hidden vectors between users and items quickly.Different from the neural collaborative filtering,the implicit vectors between users and items are embedded in two dimensions to generate feature interaction graphs as input of the neural network.Compared with simple operations,e.g.connection vectors,feature interaction graphs are capable of carrying more high-dimensional information.Then,this model employs convolutional neural network to extracting high-dimensional potential features between users and items.Owing to the parameter sharing mechanism of convolutional neural network and the powerful high-dimensional potential feature extraction ability,The performance of the convolutional neural collaborative filtering model based on denoising Auto-encoder is slightly better than the neural collaborative filtering model based on the variational auto-encoder when the convergence time of the model is greatly reduced.3)Finally,the last part of thesis compared the two models and the mainstream implicit feedback-oriented recommendation algorithms in based on two public dataset,which verifies the advantage of both models under the two indicators of Precision and Normalized Discounted Cumulative Gain.Moreover,the interval range of the optimal negative sampling number is determined through experiment.