Research On Collaborative Recommendation Algorithm Based On Bilateral Autoencoder

In the information age,serious information overload is caused by the rapid development of Internet technology.Recommendation system is an effective way to alleviate the phenomenon of information overload.By analyzing the characteristics of users and items,it helps users to select content that may be of interest from the massive information.The useritem rating matrix records the interaction information between users and items,and can be used as an information source for extracting user and item features.In the recommendation algorithm based on bilateral autoencoders,the two autoencoders work independently,resulting in the same feature will have different representations after being processed by the two different autoencoders,and the final feature interaction does not have the correction function,thus the performance of the algorithm largely depends on whether the results of the feature extraction of the two autoencoder are consistent;at the same time,the basic autoencoder does not consider the probability distribution characteristics of the samples and does not have the ability to generate,so the reconstructed samples may not have practical significance.Based on the above,an improved collaborative recommendation algorithm based on bilateral autoencoders is proposed in this paper,which takes the user-item rating matrix as the input information and the completed rating matrix as the output information.(1)A collaborative recommendation algorithm based on the unification of features of bilateral autoencoders is proposed.The multilayer perceptron is used as the feature unified transformation module of the model,so that the feature vector extracted from the encoder is transformed and then the feature interaction is carried out,which improves the cooperation between feature extraction and feature interaction.After experiments,the accuracy of the proposed model has been improved.In addition,in order to deeply study the role of each module,this paper decomposes the proposed collaborative recommendation algorithm model based on the unified interaction of features of bilateral autoencoders into"feature extraction" and "feature unified interaction" modules,and through several sets of comparative experiments to verify the performance of each module necessity.(2)A recommendation algorithm based on variational autoencoder is proposed.Considering that the variational autoencoder adds feature distribution constraints in the encoding process,so it has better generation ability and stronger robustness,this paper considers the variational autoencoder to replace the basic autoencoder in the model,and proposes A recommendation algorithm based on variational autoencoders.Experiments show that models using variational autoencoders have smaller errors.