Research And Application Of Hybrid Collaborative Filtering Recommendation Algorithms Based On Deep Learning

The rapid development of Internet technology has greatly changed the lives of the people.At the same time,it has also brought about a surge in data scale,which has increased the difficulty for users to find target information accurately and quickly.Search engines alleviate the problem of ‘information overload' to some extent,but they still cannot meet the ever-changing needs of the public.Therefore,the recommendation system is generated.Recommendation system has been widely used in all walks of life.A good recommendation algorithm can greatly improve the income of enterprises and enhance the satisfaction of users.However,existing recommendation algorithms suffer from problems such as data sparsity,cold start,recommended interpretability and other problems,resulting in low recommendation accuracy and poor user experience.With the vigorous development of deep learning technology,it has become a new trend to combine the recommendation algorithm with deep learning technology to solve the problems above.Aiming at the problem of cold start and data sparsity,the thesis comprehensively analyses the shortcomings of existing research,and tries to combine deep learning with collaborative filtering for hybrid recommendation.We make full use of side information such as item attributes and comment texts to alleviate the degradation of accuracy caused by the above problems.The main research work of this thesis is as follows:(1)A deep hybrid collaborative filtering recommendation algorithm for complete cold start is proposed.The algorithm designs a general framework that combines machine learning with collaborative filtering.In order to fully exploit the item properties,we used deep denoising autoencoder to learn latent factors of the items,and replaced the item factor.For the cold start feature,based on the learned latent factors,a safe S4 VM algorithm was used to preliminarily predict the ratings of new items.Combining with the improved LFM model considering time factor and actual situation,the prediction rating was finally generated.Theoretical analysis and experiments show that the algorithm can effectively solve the cold start problem of new items and improve the interpretability of the recommendation.(2)Aiming at the problem of data sparsity,a parallel network recommendation algorithm based on attention mechanism is proposed through the existing deep cooperative neural network.The algorithm considered word order,and used two parallel convolutional neural networks to simultaneously model the user's comment text and item description information.Aiming at the importance of words in long text,a pooling algorithm based on attention mechanism was designed and applied to parallel CNNs.Finally,two parallel networks were associated through a shared structure,and the factorization machine was used to predict the final ratings.The experimental results show that the problem of data sparsity can be alleviated by mining comment text and description information to some extent.MAE and RMSE errors are about 12% lower than the classical algorithms.(3)A prototype system of e-commerce is designed and implemented.The two algorithms above are encapsulated into a general recommendation module,which is applied in the form of popular micro-services.At the same time,considering the actual situation of the system,similar recommendation is used to improve the user's experience and satisfaction.