TY - GEN
T1 - Context-Aware Personalized Recommendation Using Large Language Models with Feedback-driven Adaptive Score Calibration
AU - Li, Yuanyuan
AU - Liu, Xin
AU - Liu, Minxia
AU - Xie, Jian
AU - Xiang, Xi
N1 - Publisher Copyright:
© 2026 Copyright held by the owner/author(s).
PY - 2026/9/16
Y1 - 2026/9/16
N2 - LLM-based recommendation methods lack effective context modeling and lightweight online feedback adaptation, failing to capture dynamic user preferences. This paper proposes a context-aware recommendation framework with feedback-driven adaptive score calibration. We design an attention module to weight multidimensional context signals and capture context-dependent preferences. A gated fusion unit adaptively combines collaborative filtering embeddings and LLM semantic representations to alleviate data sparsity. Moreover, a lightweight calibration module updates only a single scalar parameter using real-time user feedback, avoiding the high computational cost and training instability of reinforcement learning and bandit algorithms. Experiments on MovieLens-1M, Amazon-Books and Yelp demonstrate that our model outperforms mainstream baselines in Precision, Recall and F1. Ablation studies and embedding visualization verify the complementarity of all core modules. The framework balances accuracy and efficiency and is suitable for real-time recommendation scenarios.
AB - LLM-based recommendation methods lack effective context modeling and lightweight online feedback adaptation, failing to capture dynamic user preferences. This paper proposes a context-aware recommendation framework with feedback-driven adaptive score calibration. We design an attention module to weight multidimensional context signals and capture context-dependent preferences. A gated fusion unit adaptively combines collaborative filtering embeddings and LLM semantic representations to alleviate data sparsity. Moreover, a lightweight calibration module updates only a single scalar parameter using real-time user feedback, avoiding the high computational cost and training instability of reinforcement learning and bandit algorithms. Experiments on MovieLens-1M, Amazon-Books and Yelp demonstrate that our model outperforms mainstream baselines in Precision, Recall and F1. Ablation studies and embedding visualization verify the complementarity of all core modules. The framework balances accuracy and efficiency and is suitable for real-time recommendation scenarios.
KW - Adaptive Score Calibration
KW - Context-Aware Modeling
KW - Gated Feature Fusion
KW - Large Language Models
KW - Recommender Systems
UR - https://www.scopus.com/pages/publications/105052190620
U2 - 10.1145/3829918.3830005
DO - 10.1145/3829918.3830005
M3 - Conference contribution
AN - SCOPUS:105052190620
T3 - Proceedings of the 2nd International Conference on Artificial Intelligence and Product Design, AIPD 2026
SP - 570
EP - 579
BT - Proceedings of the 2nd International Conference on Artificial Intelligence and Product Design, AIPD 2026
PB - Association for Computing Machinery, Inc
T2 - 2nd International Conference on Artificial Intelligence and Product Design, AIPD 2026
Y2 - 24 April 2026 through 26 April 2026
ER -