Study On Hybrid Neural Collaborative Filtering Algorithm

With the impressive development of the informatization construction,data scale is greatly exploded,and which caused various information overload problems to interference the Internet lifes.In order to relieve these problems,the recommendation system has been widely researched and applied,related state-of-the-art algorithms have been commercialized,which are important to Internet lifes.However,traditional recommendation algorithms cannot achieve satisfactory performance under large-scale data,which have caused many problems,such as data sparseness,cold start,user interest transfer,and they are difficult in modeling implicit feedback models.In recent years,neural networks have achieved impressive results in many different research fields.Benefit from self-learning and free from designing features manually,the neural network-based methods are always discriminative and robust.Therefore,in order to relieve the existing problems,deep learning is utilized to recommend in this dissertation,which improves the performance in combining multi-features.The main contributions of this dissertations are as follows:1.In order to relieve the problem of the strong noise in the implicit feedback data,a three-channel neural collaborative-based neural collaborative filtering algorithm is proposed in this dissertation,which applies an autoencoder to users and projects.The auxiliary information is processed to extract the feature vectors of the user and the item,respectively.After that,the user and the item are modeled together with the neural collaborative filtering algorithm.In this way,the auxiliary information of the users and the items is applied to mine relatively reliable feature information between the users and the items.Meanwhile,this method can alleviate the impact of implicit feedback data noise,and it also can mine complex interactions of the relationship between the users and the items and improve the accuracy and sorting performance of the algorithm.At the same time,aiming at the negative sample lack problem,a generalized matrix factorization-based probability negative sampling method is proposed.Sufficient experiments have proofed the effectiveness of the algorithm.2.In order to relieve the problems of user interest transfer and cold start,a neural collaborative filtering algorithm integrating time-weighted is proposed.The algorithm combines the time-weighted collaborative and the neural collaborative filtering algorithm to complete the recommendation task.The advantage of this is that the using of neural collaborative filtering algorithm can relieve the data sparseness problem,which can improve the accuracy and robustness.At the same time,the time-weighted-based collaborative filtering algorithm is proposed to solve the problems of user interest changes and cold start.Sufficient experiment results have shown that the mixed algorithm can effectively improve the recommendation performance and achieve remarkable performances in different evaluation indices.