Research On Recommendation Algorithm Based On Denoising Autoencoder Model

With the development of information technology,the Internet has become an indispensable part of our daily life.However,while the Internet brought us convenience,it also brought the problem of information overload.To reduce the negative effects of information overload,the researchers developed personalized recommendation to help users filter complex information.Nowadays personalized recommendation has been widely used in business,social,tourism and other fields.Researchers found that recommendation systems combined with neural network models could effectively improve the information processing ability of recommendation algorithm.However,the recommendation algorithm combined with neural network model puts forward higher requirements on model performance,which poses a great challenge to the design of the algorithm.As for the Top-N recommendation algorithm based on the denoising autoencoder model among the various recommendation algorithms,it only uses the users' scores as the basis for judging the users' preferences when predicting the results,but does not consider the other factors affecting users' interests comprehensively.As well as when the denoising autoencoder model trains data,it ignores the effects of the randomness of the denoising for the feature expression of the data matrix,thus has an impact on the accuracy of the prediction results.For these problems,the solutions are designed respectively:Firstly,in order to solve the problem of incomplete consideration of user interest preference in the rating matrix of Top-N recommendation,a new user interest measurement standard is proposed.In the pretreatment stage,the user's specific rating value is seemed as a supplement of rating matrix,it amends the interestingness from user evaluating,then the improved matrix will be a new input matrix in the Top-N recommended training.The scheme can increase the number of correlation in the original matrix,so as to alleviate the influence of data sparseness to recommendation results.Secondly,in order to solve the problem of uneven noise distribution of the denoising autoencoder model,a multiple denoising antoencoder for Top-N recommendation is proposed.Compared with the classical autoencoder model,the improved model simulates the single noise-adding operation of autoencoder model with multiple low rates of noise-adding operations.Then,the results obtained by noise-adding of different levels are combined with a result without noise-adding,it can reduce the impact of random Gaussian noise on the integrity of the original matrix feature expression on the premise of ensuring the robustness of the acquired data.And in order to verify the effectiveness and expansibility of the scheme,a matrix decomposition model with improved Dropout layer is proposed,it can more fully retain the characteristic representation of the data matrix itself.The experimental results show that the improved scoring mechanism proposed in this thesis has certain advantages in improving the prediction accuracy of Top-N recommendation.The improved denoising autoencoder model and the improved matrix decomposition model have improved the convergence speed and prediction accuracy of the model.