VisCom ‐Res: An Efficient Visual Complementary E‐Commerce Product Recommendations Using Attention‐Enhanced Lightweight Architecture

ABSTRACT The rapid growth of e‐commerce across the world has redefined how users discover, evaluate, and purchase products especially in visually driven domains like fashion and lifestyle. While the traditional product recommendation engines mostly rely on the collaborative or textual content, and many e‐commerce platforms, providing the similar visual recommendations and lacks in suggesting the complementary products. The existing CNN baseline models overlook the necessary visual features and contextual relation which are essentials for recommending complementary products. This paper introduces VisCom‐Res which is a Pioneering framework using deep learning‐based visual complementary product recommendation combining a customized CNN architecture, ResNet50‐ Efficient Ghost Attention (EGA), with Hierarchical Navigable Small World (HNSW) nearest neighbor indexing technique. The ResNet50‐EGA is improved with Ghost modules and Efficient Channel Attention (ECA) to extract semantically rich and fine‐grained features from fashion product images. This extracted image embedding's further indexed using HNSW to enable fast, high‐precision retrieval of visually compatible items. The Proposed model evaluated on a custom curated multi‐class fashion dataset sourced from various Indian e‐commerce platforms like Flipkart, Myntra, and Tata CLiQ, and it outperforms conventional models such as VGG16, DenseNet121, and MobileNetV2 in terms of classification accuracy and achieves superior retrieval metrics such as Precision@10 (93.8%), Recall@10 (94.5%), NDCG@10 (0.947), and average query time of 8.2 ms. The results validate the suitability of VisCom‐Res for scalable, real‐time recommendation systems in visually rich e‐commerce recommendation environments. Unlike conventional CNN + attention frameworks, VisCom‐Res integrates Ghost feature generation with Efficient Channel Attention (ECA) within a ResNet‐50 backbone, reducing redundant map computation by ≈30% while retaining complementary‐aware semantics.