Bundle Recommendation with Attribute-Level Preference Elicitation and Personalized Attention for Balancing Accuracy and Diversity

Abstract

Increasing user exposure through highly personalized digital services remains a crucial element for modern e-commerce growth. Contemporary customer consumption patterns involving purchasing multiple items simultaneously have driven the rapid emergence and widespread adoption of Bundle Recommender Systems (BRS). However, traditional approaches utilizing latent feature-based methods, such as Deep Learning and Matrix Factorization successfully capture hidden user-item interaction patterns, yet they fundamentally lack transparency and interpretability. Conversely, traditional feature engineering approaches offer improved interpretability, but their overall real-world effectiveness is heavily constrained by intensive manual tuning processes and overly simplistic feature representations. To resolve these critical challenges, this study introduces an innovative framework named b-ANDY (bundle recommendation with Attribute-level iNterpretation for Diversity and relevancY). The proposed b-ANDY framework directly incorporates powerful Large Language Models (LLMs) into automated feature engineering pipelines, encompassing systematic feature extraction, feature transformation, feature selection, and feature creation. This novel architecture optimizes overall recommendation quality by carefully balancing item diversity with consumer relevance, while simultaneously providing highly interpretable, user-friendly explanations for end users. Extensive empirical evaluations demonstrate that the b-ANDY framework successfully attains an impressive absolute diversity score (Inter-Diversity@10) reaching 61.2% alongside a robust relevance score (Hit Ratio@20) of 47.5% across modern recommendation environments. Consequently, this study successfully delivers a scalable, balanced, and exceptionally reliable solution specifically designed for next-generation e-commerce platforms.