B2-PFR: Deep bidirectional learning of user-item body features for fashion compatibility modeling

State-of-the-art (SOTA) Personalized Fashion Recommender (PFR) systems increasingly integrate user behavior, item semantics, and body shape attributes. These body-related features are typically derived from anthropometric data (e.g., height, bust, waist). However, in real-world scenarios, such body measurements are often unobtainable, diminishing model efficacy. To tackle this challenge, we propose B2-PFR, a novel PFR method rooted in SOTA frameworks. It innovatively integrates image-based user body features with item-target body information, using anthropometric data as auxiliary input, to deeply model user-item compatibility and boost recommendation performance. Specifically, to accurately capture user body shapes, we extract global body characteristics from 3D Skinned Multi-Person Linear (SMPL) models (built with optional anthropometric data) and detailed body proportions from their multi-view 2D projections. Moreover, to explore latent item-target body suitability, we derive item-target body features by mining body shapes from implicit user-item interactions and fusing them with learnable attention weights, enabling adaptive prioritization of context-relevant body traits. User preference embeddings are further optimized by attention-based aggregation of visual-textual features from paired items to model genuine inclinations. Additionally, we construct the Dress4BodyShape dataset, curated to recommend dresses for various body shapes. Evaluations on this dataset show that B2-PFR outperforms SOTA recommenders in recommending preference-aligned and body-aware dresses, achieving 7.2% higher Recall@10 and 5.2% higher NDCG@10 than the best baseline. It also achieves more accurate preference understanding on the IQON3000 dataset and effective body shape modeling on the 2D image-to-BMI dataset.