Research On Explainable Fashion Recommendation Combined With Comment Generation

Fashion is playing a more important role in our society because of its ability to show chara.cter and build culture.In recent years,the online ret,ail market of fashion products such as clothing,shoes,and jewelry has continued to prosper,reflecting the huge demand for fashion goods by the public.Everyone wants to have the charm of fashion.but not everyone is an expert in fashion.In the face of a dazzling array of fashion goods in the online retail market,how to choose and match is a big problem for most people.From the perspective of the merchant,how to recommend afashionable outfit with credibility to the user is also crucial to increase the sales volume of the product.The fashion recommendation technology is to solve the above problems.At the same time,in recent years,a large number of fashion-oriented communities have emerged.Fashion experts can share their fashion secrets in the community,And fashion recommondation technology can acquirerich data from these communities to improve the current recommendation performance.This article aims to address the outfit matching problem in fashion recommen-dation,i.e.recommend matching bottoms(such as jeans,skirts,etc.)for a given top(such as T-shirt,jacket,etc.)and vice versa.Most of the previous work on fashion recommendation focused on designing or extracting effective visual fea-tures to enhance the recommendation performance.Existing work ignores user comments for fashion outfits.In the recommendation system for other fields,these comments have proven to be very useful for generating explainable and better recommendation results.This paper proposes a novel fashion recommendation framework based on deep learning,the Explainable Outfit Recommender(EOR),which can simulate users to generate comments while recommending a fashionable outfit.The generated comments can express the public opinion t.o the recommended outfit,which can be regarded as the explanation to increase the credibility of the recommendation.The framework consists of two parts:an outfit matching part and a comment generation part.For the outfit matching part,this paper first uses a convolu-tional neural network to extract the visual features from the image of a top and the image of a bottom;then the paper proposes a mutual attention mechanis-m to encode the matching relation between the top and the bottom into their visual representations;finally,the paper uses a multi-layer feed-forward neural network to decode their visual representations into a rating score for the match-ing prediction.For the comment generation part,this paper proposes a gated recurrent unit network with a cross-modality attention mechanism to transform visual features into a concise sentence as the comment and the explanation.The cross-modality attention mechanism can help our network use visual informa-tion to generate text more efficiently.These two parts can be trained jointly by a multi-task deep-learning framework based on an end-to-end back-propagation algorithm.This article collects a large dataset from an online fashion-oriented community,which contains a large number of user comments for outfits.And the paper also uses a small existing dataset.Extensive experiments conducted on these two datasets show that the proposed framework achieves the significant improvement on MAP,MRR and AUC compared to the state-of-the-art baselines.And the generated comments achieve impressive ROUGE and BLEU scores in comparison to real comments written by humans.The generated comments can be considered as the explanation to the recommendation results.In addition,the paper has validated the effectiveness of the two attention mechanisms through a series of experiments.