Partially Supervised Compatibility Modeling.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fashion Compatibility Modeling (FCM) research increasingly focuses on disentangling item representations.
• Existing FCM methods primarily utilize visual content, neglecting semantic attributes like color and pattern, which limits performance and interpretability.

Purpose of the Study:

• To develop a comprehensive approach for Fashion Compatibility Modeling (FCM) by integrating both visual content and semantic attributes of fashion items.
• To address challenges in partially supervising attribute-level representation learning, ensuring disentangled representations, and effectively combining multi-granularity information for improved performance and interpretability.

Main Methods:

• Proposed a partially supervised outfit compatibility modeling scheme (PS-OCM).
• Devised a partially supervised attribute-level embedding learning component to disentangle fine-grained attribute embeddings.
• Introduced a disentangled completeness regularizer to prevent information loss during disentanglement.
• Designed a hierarchical graph convolutional network to integrate attribute- and item-level compatibility modeling for explainable reasoning.

Main Results:

• The proposed PS-OCM significantly outperforms state-of-the-art baselines on a real-world dataset.
• The method effectively disentangles attribute-level representations and integrates them with item-level information.
• The hierarchical graph convolutional network enables explainable compatibility reasoning.

Conclusions:

• The PS-OCM offers a significant advancement in Fashion Compatibility Modeling by incorporating semantic attributes.
• The approach enhances both the performance and interpretability of outfit compatibility prediction.
• The study provides valuable insights and resources for future research in fashion AI.