Learning to represent causality in recommender systems driven by large language models (LLMs)

• Published in: Discover applied sciences Vol. 7; no. 9; pp. 960 - 27
• Main Authors: Aman, Serge Stéphane, Kone, Tiemoman, N’guessan, Behou Gerald, Kimou, Kouadio Prosper
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.09.2025; Springer Nature B.V; Springer
• Subjects: Algorithms; Applied and Technical Physics; Artificial intelligence; Bayesian analysis; Bayesian network; Causal representation; Causality; Chemistry/Food Science; Collaboration; Data processing; Decision making; Deep learning; Earth Sciences; Engineering; Environment; Error analysis; Language; Large language models; Large language models (LLMs); Machine learning; Materials Science; Mathematical models; Natural language; Probability; Product reviews; Recommender systems; Semantics; User behavior; Variables; Causal representation; Causal inference; Large language models (LLMs); Machine learning; Reinforcement learning; Personalized recommendation; Natural language processing (NLP); Bayesian network; Artificial intelligence; Recommender systems
• DOI: 10.1007/s42452-025-07551-8
• ISSN: 3004-9261; 2523-3963

Current recommender systems mainly rely on correlation-based models, which limit their ability to uncover true causal relationships between user preferences and item suggestions. In this paper, we propose a hybrid model that combines a Bayesian network with a large language model (LLM) to enhance both the relevance and interpretability of recommendations. The Bayesian network captures causal dependencies among user-item interactions, while the LLM injects contextual semantics from user reviews and product descriptions. Our method was evaluated on a dataset of 1.2 million interactions and showed significant improvements over baseline models, with gains of 84.44% in precision, 88.37% in recall, and 89.36% in NDCG. A statistical t-test confirmed the significance of these improvements ( p  < 0.05). We further provide an error analysis and discuss the implications of using causal modeling for scalable, transparent, and GDPR-compliant recommender systems. Our results underscore the potential of causal representation learning to improve personalization and decision-making in recommender systems.