Research And Application Of Commodity Recommendation Algorithm Based On Deep Learning Model

In the era of big data information technology,online shopping provides people with convenience,but the variety of products makes it difficult for users to choose the products they need.Recommendation systems can effectively alleviate the phenomenon of ”infor-mation overload” and help users quickly locate products that may be of interest to them.In recent years,deep learning has been widely used in various fields,and recommendation algorithms based on deep learning are favored by researchers.However,recommendation algorithms still suffer from data sparsity,insufficient use of implicit feedback,changing user interests,loss of sequence information,and difficulty in scaling deep models.To ad-dress the above problems,this thesis studies the current recommendation algorithms and combines their shortcomings to propose two solutions using the deep learning framework.The main research contents of this thesis are as follows.(1)Analyzes the principles of collaborative filtering algorithms and word embedding techniques,and proposes UI2 vec,a collaborative filtering algorithm based on embedding representation,for the problems that collaborative filtering algorithms such as matrix de-composition have significantly reduced performance when data is sparse and implicit feed-back is underutilized.According to the joint feature extraction network designed in this section,UI2 vec embeds users and items on the potential space at the same time,and uses the item similarity between them to predict the user's content of interest.Then a genera-tive model VUI2 vec with more stable performance is proposed based on UI2 vec,which maps users and items as independent Gaussian distributions and obtains the approximate posterior distribution of both by variational inference.(2)GAT4Rec,a sequence recommendation model based on self-attentive mask learn-ing,is proposed for the problems of user interest change,sequence information loss,and difficulty in extending the depth model.the method designs a gated filtering layer for calculating the similarity of user preferences,makes full use of the edge information of items,and extracts the interaction sequences as user interest-related subsequences.Posi-tion embedding is added to retain sequence information,and the extracted subsequences are transformed into user-item deep feature combinations by a self-attention-based coding layer with a parameter sharing mechanism.Finally,a joint training loss function combined with mask learning is designed.(3)The recommendation performance of UI2 vec and VUI2 vec is evaluated on Ta Feng and Movielens datasets? Movielens and Taobao datasets are chosen to verify the effective-ness of GAT4 Rec.The impact of important superparameters within the model on perfor-mance is investigated,and visualization experiments are conducted to better understand the principles of the model.The experimental results show that compared with the base-line model optimum,the metrics F1-score of UI2 vec and VUI2 vec increase by 6.80% and8.37% on the Ta Feng dataset and by 2.47% and 3.91% on Movielens,respectively? the experimental metrics NDCG@10 of GAT4 Rec on the four datasets increase by increased by 5.77%,1.35%,11.58%,and 1.79%.(4)Combining the two proposed recommendation algorithms,a product recommen-dation system was designed and implemented based on B/S architecture.From require-ment analysis to system design the display layer,logic layer,and data layer are planned to be responsible for,and in the recommendation process GAT4 Rec and VUI2 vec rec-ommendation algorithms are applied to the recall and sorting stages of the commodity candidate set,respectively,and finally the system is implemented based on Springboot,Vue,Flask,and PyTorch.