Modeling anisotropic preference manifolds for robust graph-based fashion recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph Convolutional Networks (GNNs) are prevalent in collaborative filtering but struggle with complex fashion data due to geometric mismatches.
• Traditional methods assume isotropic user interests, failing to represent the anisotropic variance in fashion preferences.

Purpose of the Study:

• To address the limitations of existing GNNs in fashion recommendation by proposing a novel framework.
• To develop a model that accurately captures complex and diverse user preferences in fashion.

Main Methods:

• Proposed the Multi-Interest Mahalanobis Denoising Graph Convolutional Network (MIMD-GCN).
• Introduced a Poly-Attention mechanism for disentangling user representations into multiple latent interest centers.
• Developed an anisotropic denoising module using a learnable Mahalanobis distance for geometry-aware denoising.

Main Results:

• MIMD-GCN demonstrated consistent improvements in recommendation performance on Amazon-Clothing and Taobao datasets.
• The proposed model showed enhanced robustness against synthetic noise compared to baseline methods.
• Anisotropic modeling proved beneficial for capturing intricate user preferences in fashion.

Conclusions:

• The MIMD-GCN framework effectively models complex user preferences in fashion recommendation by addressing geometric mismatches.
• Anisotropic modeling and geometry-aware denoising are crucial for enhancing GNN performance in specialized domains.
• The study highlights the importance of considering the geometric structure of data for improved recommendation systems.