An Analytical Model using CVAE-based Image Generation from Product Descriptions and Image Data
Fuchi Ayuno, Ayako Yamagiwa, Masayuki Goto · Industrial Engineering & Management Systems · 2025
Customers who chooseproducts on EC sites refer to product attributes, images, and descriptions. The impressions given by images are critical because visual information has a significant impact on customers’ purchase decisions. It is, therefore, important to analyze the relationship between customers’ needs and product images, so several studies to analyze product images from product attributes, invoking deep learning models, have been reported recently. For ex- ample, the method can analyze product images based on customers’ and products’ attributes. However, the abstract needs of customers are not always reflected by only product attributes. On the other hand, some words in product de- scriptions express abstract images of products to express customers’ abstract needs. Therefore, if we can model the relationship between each product by using text information, such as product descriptions, and image information, such as product images, it is possible to capture the product line-up from a higher perspective. In this study, we pro- pose a product image analysis method that uses latent topics extracted from product descriptions using Latent Dirichlet Allocation as conditions of Conditional Variational Autoencoder. The topics reflect product image related to customers’ needs. Applying the proposed method to the actual dataset, we show that it enables capturing product features based on abstract information extracted from product descriptions, more than just considering each feature separately