Research On Personalized Recommendation Algorithm Based On Convolutional Neural Network And Attention Mechanism

With the rapid development of artificial intelligence technology and the explosive growth of network data scale,users cannot quickly obtain their favorite items,resulting in a poor user experience.As an important method that can effectively alleviate "information overload",the personalized recommendation algorithm can help users accurately locate the items they need.In recent years,deep learning has been successfully applied in various fields of society,making individualized recommendation algorithms based on deep learning favored by researchers in the field of recommendation.Compared with traditional recommendation algorithms,the personalized recommendation algorithms based on deep learning have a multi-layer nonlinear structure,which can more effectively mine the deeplevel representation of features.However,the individualized recommendation algorithm based on deep learning still has problems such as difficulty in obtaining effective crossfeatures,low feature utilization,difficulty in expressing user interests,poor algorithm interpretability,and weak generalization ability.Given the above-mentioned problems and the deficiencies of other latest personalized recommendation algorithms,the thesis proposes two personalized recommendation algorithms based on the deep learning framework.The main research contents of this article are as follows:1.Aiming at the difficulties in obtaining effective cross-features,low feature utilization,and too complex algorithm structure in current personalized recommendation algorithms,a self-attention recommendation algorithm(MSRN)with low-complexity feature automatic association is proposed.This algorithm uses the idea of convolutional neural network convolution kernel up and down convolution and parameter sharing,to achieve the goal of automatically obtaining effective low-order cross feature matrix under the condition of less algorithm parameters,and then with the high-order cross feature obtained by the deep neural network layer The matrix is input to the self-attention mechanism layer,and the cross features are weighted and summed,and different weight values are assigned to the cross features,so as to improve the feature utilization rate of the algorithm.2.Aiming at the problems of difficulty in expressing user interests,poor interpretability and weak generalization ability in current personalized recommendation algorithms,a multi-head self-attention recommendation algorithm(IARM)with interpretable feature combination was proposed.The algorithm first uses user information to form user interest representations from multiple perspectives,and then designs an interaction layer that can increase the interpretability and generalization capabilities of the algorithm.In addition,the interaction layer can also realize the automatic acquisition of low-order and high-order effective intersection features on the basis of preserving the original information of the features.3.The recommendation performance of MSRN and IARM is evaluated on three classic recommendation algorithm datasets,Criteo,Avazu,and Movielens,and the effects of algorithm hyperparameters and algorithm structure on the algorithm performance are studied.At the same time,the automatic acquisition of the algorithm’s effectiveness features is verified and the algorithm is visualized.The experimental results show that compared with other state-of-the-art recommendation algorithms,both MSRN and IARM have better performance on three public datasets,and have higher recommendation accuracy,which verifies that the algorithm has excellent interpretability and generalization.