Research On Technologies And Methods Of Recommendation System Based On Neural Embedding

With the rapid development of information technology and the Internet,the whole world from the lack of information to the era of information overload.It becomes more difficult for users to find their interesting content accurately and quickly from the mass of information.To address this problem,a personalized recommender system is applied.Its construction mainly consists of three methods:content-based recommendation algorithm,collaborative filtering recommendation algorithm,and hybrid model recommendation algorithm.The main propose of Collaborative filtering algorithm is to analyze the user’s historical interaction behavior(e.g.,click and purchase)and provide a personalized recommendation service.Matrix factorization is widely used in various collaborative filtering technologies.This method uses latent feature vectors to represent users or items and project them into the shared latent vector space.Compared with the matrix factorization model,which has only two types of features(users and items),the factorization machine model considers more features.The initialization of matrix factorization and factorization machine is mostly random.This does not make full use of the rating matrix(e.g.,click and buy)data.According to the characteristics of matrix factorization and factorization machine algorithm,this thesis focuses on the following four aspects:·Aiming at the problem of data sparseness,word vectors are applied to matrix factorization,and a generalization model of matrix factorization based on neural embedding is proposed.First,a probabilistic auto-encoder is used to generate the neural embedding vector of the item from the user-item data.After that,the item vector is combined with the regression model based on single point negative sampling,and the regression coefficient is used to represent the hidden feature vector of the user.Finally,the inner product is applied to the latent characteristics of the user and the item to determine the correlation between them.It should be pointed out that the proposed method is generic,and matrix decomposition can be expressed and promoted under its framework.In this study,we apply ridge regression learning to the latent feature vector of each user.Experimental results on two benchmark data sets show that this model is superior to other latest methods.·Aiming at the problem that it is difficult to simultaneously obtain the local and global correlation for factorization machine,an effective factorization machine model based on probabilistic auto-encoder is proposed.The traditional factorization machine model has poor performance in simultaneously capturing the local and global structure of user-item correlation.Although deep neural networks have been used to improve factorization machines,deep networks have increased the complexity of the training process.A factorization machine model method based on probabilistic auto-encoder(AutoFM)is proposed to solve this problem.This method extracts non-trivial local structural features from user-user/item-item co-occurrence pairs by integrating a low-complexity probabilistic auto-encoder.Moreover,it supports explicit and implicit feedback data sets.Extensive experiments on four real-world data sets demonstrate the effectiveness of AutoFM.The results show that the proposed method is superior to the existing techniques in rating prediction tasks.Compared with the model based on deep neural networks,the proposed model’s ranking is improved by at least 1.16%～4.3%.·Aiming at the random initialization problem of negative sampling deviation and Singular Value Decomposition(SVD),neural network embedding SVD(NESVD)is proposed for collaborative filtering.NESVD use item popularity as a weighting factor,and neural embedding as the initial singular value decomposition and use it for collaborative filtering.Singular value decomposition is one of the most effective algorithms in recommender system.Due to the iterative nature of the singular value decomposition algorithm,a big challenge is initialization,which has a great impact on the convergence and performance of the recommendation system.Unfortunately,the existing singular value decomposition algorithms usually initialize the user and item features in a random manner,so they do not make full use of the user-item data information.Aiming at how to develop an effective singular value decomposition algorithm initialization method.A general neural embedding initialization framework is proposed,which uses a low-complexity probabilistic autoencoder neural network to initialize the features of users and items.The framework supports both explicit and implicit feedback data sets.The design details of our proposed framework are elaborated and discussed in detail.The experimental results show that the initialization framework of recommender system based on nerual embedding is improved by at least 2.20%～5.74%compared with the existing algorithms and other matrix factorization methods in the literature.·Aiming at the problem of random initialization of factorization machine,a neuralembedded factorization machine model is proposed for user response prediction.Since factorization machine models only describe the interaction between features linearly,they cannot accurately capture the non-linear and complex relationships of the data.In addition,the random initialization in the factor machine model seriously affects the convergence and performance of the system.Moreover,the factorization machine model cannot make full use of data information.Although models based on deep neural networks have recently been proposed for advanced feature interaction,training deep structures are complex.To solve these problems,a neural embedding factorization machine model is proposed,which is based on the unsupervised pre-training framework of probabilistic auto-encoders,which effectively initializes the embedding layer.The proposed method skillfully combines the good linearity of the factorization machine model in second-order feature interaction modeling and the advantages of deep network structure in modeling nonlinear feature interaction.Experimental results show the effectiveness of the method.For example,the proposed method’s performance is at least 6.99%higher than that of the random initialization model.Compared with the pre-training model based on deep networks,the proposed method’s test RMSE is improved by at least 1.02%.