Product Representation Learning Using Images and Text Features for ecommerce

Online e-commerce shopping platforms contain a very huge collection of products, and users often find it difficult to identify the most relevant items from such huge collections. For this reason, product representation plays an important role in building effective and efficient recommendation systems. Many traditional approaches depend mainly on user ratings, purchase history, or a single type of product information. However, a product is better need to understand when both its image and related details such as name, category, colour, usage, and description are considered together.

This paper presents a simple approach for learning product representations using product images and metadata. In this approach, image-based features are extracted from product pictures, while metadata-based features are generated from available product information. These features are then combined to create a unified representation for each product. The generated representation is used to measure product similarity and retrieve related products. The approach is evaluated using Retrieval@5, Retrieval@10, and NDCG@10. The results show that combining product images and metadata provides better product representation than using only image-based or metadata-based features. This work provides a basic foundation for improving product similarity learning in e-commerce recommendation systems.