Research On Clothng Matching And Recommendation Based On Self-attention

People will be able to automatically attain satisfactory clothing matching as intelligent technology progresses.Deep learning-based technology for matching and also recommending clothing is increasingly coming to the fore.However,a few problems exist today.On the one hand,the clothing matching model used in earlier studies has limited capacity for processing and analyzing clothing text description data;on the other hand,it is too simple for the fusion processing of top and bottom clothing features(including image and text description data).Additionally,the modeling of a user’s liking for one specific piece of clothing in the personalized clothing recommendation model from earlier research is very simplistic,making it impossible to adequately investigate the relationship between users and clothing.And the user’s preference for an outfit is not taken into account when providing recommendations for them.This study creates a three-stage neural network architecture for a novel clothing matching model by designing and adding a feature fusion structure based on the self-attention mechanism to address the aforementioned issues.Then,A new personalized clothing recommendation model is built based on this model,which also serves as the foundation for further designing the user’s preference model for clothing.This recommendation model fully takes into account the user’s personal preference for an outfit in addition to the "expert" advice(provided by the previous clothing matching model)and the user’s own preference for a specific piece of clothing.In the clothing matching model proposed in this study:(1)To acquire more thorough and useful text features,the feature extraction model of clothing text description is improved,and the bidirectional long-short-term memory neural network(Bi-LSTM)is employed to process the text description information of clothes.The AUC of the model utilizing the Bi-LSTM network was 0.693,an increase of 1.1% when compared to the prior model GBPR.(2)Design and add a fusion structure for clothing features based on the self-attention mechanism.According to the test,the AUC hits 0.741 after employing this structure,5.9% higher than GBPR.As a result,it is simpler for the model to create the relationship between clothing features and clothing matching attributable to the clothing fusion structure based on the selfattention mechanism suggested in this study.The new clothing matching model that is suggested in this paper can achieve an AUC of0.750.The evaluation index has changed significantly from the prior model.Along with the indicators being better,the results of various application scenario experiments are also getting better.In the personalized clothing recommendation model proposed in this study:(1)The user’s preference model for one piece of clothing is optimized through the use of neural collaborative filtering algorithm(NCF).The model can extensively mine the relationship between users and clothing mainly to the formation of fully connected layers.(2)To increase the user’s contentment with the customized requirements of the recommendation results,design and add the model of the user’s preferring relationship between the top and bottom outfit.According to the test results,the personalized clothing recommendation model provided out in this study has an evaluation AUC of 0.8559,which is 1.7% higher than the classic model GPBPR.The clothing recommendation model has improved by 10% in terms of indicators after adding the user’s clothing preference model created in this study,proving the significance of individual preferences in the task of personalized clothing recommendation and demonstrating the validity and reasonableness of the model’s appropriate design.