Image-Based Fashion Recommendation with Attention to Users' Interests
Wenxin Yao, Kaoru Uchida, Jianpeng Zou, Xiaolin Zhong · 2022
Building an effective fashion recommendation system is challenging due to the high level of subjectivity and the semantic complexity of the features involved. Users’ decision depends largely on their interest and the appearance of the product. Such information is often hidden in implicit feedback from users' purchase histories and product images. Most interest-based recommendation systems like Deep Interest Network (DIN) and Deep Interest Evolution Network (DIEN) only take advantage of product attributes and context review, which are all text information. There are also some studies focusing on the use of image features for a fashion product recommendation. They try to extract features from images and recommend products based on their similarity. However, for DIEN. It works not well when there are few interactions between users and items, the model can not find users' interests effectively. On the other hand, these image-based recommendation systems tend to ignore an important factor: the user's interest. We propose a new system trying to find users' interests by introducing the visual information of the product and our image-based deep interest attention model based on DIEN. Product attributes and user preference can both be represented by introducing visual information. We can model similar attribute items in the same place and find user interests more effectively. We conducted a series of experiments on our own dataset to compare user click-through rate (CTR) AUC and Accuracy using our model with DIEN and other existing models.Our experimental results demonstrate that visual information effectively aids CTR prediction and achieved better recommendation results.