계층형 이미지-텍스트 멀티모달 모델을 활용한 국내 스타트업-해외 바이어간 B2B 추천 성능 향상 연구

This study aims to solve the 'semantic interference' problem that occurs in B2B recommender systems when processing over 17,000 product categories within a single embedding space. To address this, we propose a two-stage Hierarchical Multimodal Recommendation Model (Hierarchical Model). The proposed model first predicts one of 238 mid-level categories in the first stage, and then performs detailed item recommendation on a refined candidate pool within that category in the second stage. Comparative experiments against a single model (Baseline Model) using the same CLIP-based architecture showed that the Hierarchical Model achieved a consistent and significant performance improvement of 4.5-5.0%p on average across all key metrics, including Precision@K, MAP@K, and NDCG@K. This paper empirically demonstrates that this performance enhancement is not merely due to search space reduction, but is based on clear theoretical justifications: (1) 'semantic disambiguation' through the specialization of embedding spaces, (2) the efficiency of a 'learned cascade' structure, and (3) the mitigation of data sparsity via the 'information bottleneck' principle.