Optimization Of Neural Collaborative Filtering For Web Pages Recommendation

The Web page is the most significant form of Internet content delivery,and the recommendation system is one of the Internet’s core application technologies.Deep neural networks have achieved tremendous success in speech recognition,computer vision,and natural language processing since the advent of deep-learning.However,deep neural networks have been less explored in the field of recommendation systems than in the aforementioned areas,and when it comes to the key factor of collaborative filtering modeling,that is,the interaction between user and item features,the current work tends to rely on matrix decomposition methods and inner product operations on the potential features of users and items,which are susceptible to the problem of missing user interaction information.This study concentrates on neural network-based techniques to investigate the key problem of optimizing collaborative filtering in implicit feedback-based web recommendation algorithms to solve this issue.The major contributions of this thesis are as follows:(1)A deep collaborative recommendation method based on a CNN with an outer product matrix and hybrid feature selection features,the introduction of a stack interaction map to increase the expressiveness of input features,and the adoption of an interaction map to encode more potential signals are the principal technical contributions of this thesis.In addition,CNN and background-based interaction map techniques for learning user-item feature relationship information are introduced.(2)A hybrid feature selection module is designed to learn local and global item correlations using point-by-point convolution,general average pooling,and atrous multiscale mechanisms to effectively capture item correlations.To prevent overfitting,the method also employs the weights of generalized matrix decomposition to optimize the overall network performance.We introduced the CNN residual module that regained loss while learning user-item relationships.The technical approach proposed in this thesis was verified through a series of experiments on multiple datasets,thereby demonstrating its efficacy.