Research On Recommendation Model Based On Deep Learning

The meaning of the recommendation system is to recommend the products or in-formation they like to the user,and take the user's interest characteristics and the user's historical behavior as the basis.As deep learning technology expands more and more in the field of natural language processing,the recommendation algorithm gradually re-lies on deep learning technology,and deep learning has gradually become an important field competing by major Internet companies.At present,there are still big problems in the field of text extraction and recommendation algorithms in deep learning,such as poor interpretability of neural network models,inability to fully model available information in feature extraction,and logic and accuracy of recommendation algorithmsNot strong enough.This thesis mainly studies the technology of deep learning in feature extraction and recommendation algorithms,and analyzes the advantages and disadvantages of these tech-nologies and its internal principle mechanism.In view of the defects of existing methods,this thesis mainly focuses on the users of the recommendation system field.Information extraction and recommendation algorithms have made related improvements,and finally verified the superiority of its effect through experiments.The main work of this article is as follows:1.Analyzed the basic principles of the mainstream methods of information extraction in the field of natural language processing such as: RNN,CNN,attention mechanism,and word embedding representation.Aiming at the problems of poor controllability and insufficient interpretability of the traditional self-attention mechanism,this thesis first de-signs an attention mechanism that uses a restriction function to extract more purposeful features.Next,this thesis uses the idea of word embedding combined with a restricted attention mechanism to design a layer of independent embedding layer networks,so that the embedding layer network can be used to train embedding vectors that are sufficient to express the user's interests.2.In this thesis,the embedding vectors trained by the embedding layer network are used to design a recommendation algorithm based on the similarity of the user candidate set to solve the problems that the baseline model recommendation algorithm is too simple and the candidate set is improperly selected.This recommendation algorithm can be better usedThe user's feature information is used to make targeted recommendations that are more in line with user preferences,and the algorithm is optimized for problems such as large model calculations and cold start.Finally,comparative experiments were carried out on the improved model and the baseline model,and it was verified that the improved model can make more purposeful recommendations,and it is better than the baseline model in terms of accuracy and recall.