Scholarlink

<p>Recommendation systems are integral to online platforms and e-commerce applications, aiding users in discovering personalized and relevant content or products. Recent advancements have showcased the effectiveness of attention mechanisms in enhancing recommendation system performance by focusing on informative data segments. However, there is a need to retrieve similar products based on the style available in the product image, which is crucial for recommendation systems. Motivated by the need to identify style-based patterns, this study leverages style attention-based deep learning architectures to enhance recommendation system performance. The proposed approach comprises two key tasks: Query Product Classification and Extracting Product Recommendations. The first task introduces a Dual Attention Network, which takes the attention weights jointly obtained from the Channel Style-Based Attention Module (CSAM) and Gated Style-Based Attention Module (GSAM) to categorize the query product. This approach starts with extracting two feature representations for the query product image using Xception and VGG-16. The channel attention is applied to the feature representation derived from the Xception, which is then processed through the Style-Based Attention Module. Similarly, gated attention is applied to the feature representation derived from VGG-16, which also passes through the Style-Based Attention Module, to capture style patterns in the local feature maps. Subsequently, these features are fused and fed into a network to categorize the query product. The cosine similarity is computed to retrieve analogous products, enabling the detection of visually similar items within the product repository. By analyzing style-related features, recommendation systems gain valuable insights into users’ underlying preferences and affinities, which may not be overtly articulated. By integrating attention mechanisms and diverse feature representations, the proposed system achieves an accuracy of 92.16%,82%, and 98.27% on the Fashion Product Images, Shoe, and Shoe vs Sandal vs Boot Image datasets, respectively, outperforming numerous existing models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fusion of pre-trained model features with style-aware attention for enhanced content-based recommendation system

  • 期刊论文

  • 发表日期：

    2026年1月25日

查看全文
  • B. Suvarna,
  • Sivadi Balakrishna

摘要

Recommendation systems are integral to online platforms and e-commerce applications, aiding users in discovering personalized and relevant content or products. Recent advancements have showcased the effectiveness of attention mechanisms in enhancing recommendation system performance by focusing on informative data segments. However, there is a need to retrieve similar products based on the style available in the product image, which is crucial for recommendation systems. Motivated by the need to identify style-based patterns, this study leverages style attention-based deep learning architectures to enhance recommendation system performance. The proposed approach comprises two key tasks: Query Product Classification and Extracting Product Recommendations. The first task introduces a Dual Attention Network, which takes the attention weights jointly obtained from the Channel Style-Based Attention Module (CSAM) and Gated Style-Based Attention Module (GSAM) to categorize the query product. This approach starts with extracting two feature representations for the query product image using Xception and VGG-16. The channel attention is applied to the feature representation derived from the Xception, which is then processed through the Style-Based Attention Module. Similarly, gated attention is applied to the feature representation derived from VGG-16, which also passes through the Style-Based Attention Module, to capture style patterns in the local feature maps. Subsequently, these features are fused and fed into a network to categorize the query product. The cosine similarity is computed to retrieve analogous products, enabling the detection of visually similar items within the product repository. By analyzing style-related features, recommendation systems gain valuable insights into users’ underlying preferences and affinities, which may not be overtly articulated. By integrating attention mechanisms and diverse feature representations, the proposed system achieves an accuracy of 92.16%,82%, and 98.27% on the Fashion Product Images, Shoe, and Shoe vs Sandal vs Boot Image datasets, respectively, outperforming numerous existing models.