Attention Based Deep Learning Approaches for Enhanced Product Recommendation Systems

Abstract

Recommendation Systems have become an integral component of online platforms and e-commerce applications, offering personalized product recommendations based on user preferences. The rapid expansion of online shopping amplified the demand for image-driven recommendation systems that emphasize visual attributes, enabling users to discover products aligned with their preferences efficiently. Unlike traditional recommendation systems that rely on descriptive customer and product details, image-based content-driven systems focus exclusively on extracting meaningful visual features to enhance recommendation accuracy. newlineTraditional statistical techniques have been shown to be effective to some extent; however, they often fail by producing irrelevant recommendations when user preferences are more visually driven rather than descriptively articulated. As the volume of data increases, efficiently extracting relevant product recommendations from large datasets becomes increasingly challenging. This research addresses these challenges by employing a two-stage process for product recommendation: Query Product Classification and Product Recommendation Retrieval. The first stage focuses on classifying the query product based on its visual features using advanced deep-learning approaches. The second stage retrieves visually similar products from the database by calculating their similarity, effectively identifying comparable items based on visual attributes. This two-stage approach ensures precise product classification and retrieval, significantly enhancing the performance of the recommendation system. newlineThe first approach for Query Product Classification proposes a \ac{dl} model that leverages pre-trained models such as Xception and VGG16 as backbones to extract high-level feature representations from the query product image. These features are processed through a Two-Channel Deep Neural Network (Two-Channel DNN) to accurately predict the category of product.