<p>In the era of information overload, recommender systems have become an integral part of our daily lives. These systems can recommend items, products or services to users based on their preferences and past behavior. Explainable recommender systems (XRS) extend this functionality by providing human-understandable justifications for the recommendations. It helps enhance user trust and system transparency. Traditional XRS models primarily relied on small transformer-based language models (SLMs). However, their explanations were often generic and repetitive. Modern XRS frameworks have employed large language models (LLMs) which generate contextually richer explanations. However, relying solely on LLMs makes the model computationally expensive and resource-intensive. To address these limitations, this paper proposes a hybrid XRS framework that combines the strengths of both SLMs and LLMs leveraging data distillation. The framework consists of three phases. In phase one, user and item profiles are constructed by extracting feature information from historical reviews. These profiles become part of the input prompt for data distillation process. In the second phase, raw user reviews are summarized into structured and generalized explanations using an LLM. These serve as the output explanations for the data distillation process. Having built the input–output explanation dataset, lightweight SLMs (GPT-2 and T5-small) are fine-tuned in the last phase. The fine-tuned models are capable of generating rich, interpretable explanations comparable to those of LLMs, with a significant reduction in computational cost. This makes them suitable for scalable and resource-constrained environments. Experiments on three real-world datasets (Yelp, Amazon Fine Foods and Amazon Video Games) demonstrate that our distilled SLMs outperform prior SLM-based frameworks (NRT, PETER, PEPLER) and perform competitively against pure LLM-based frameworks (XRec) across evaluation metrics like BERTScore, BARTScore and BLEURT. In addition, the proposed framework enables low-latency and energy-efficient explanation generation, as evidenced by a comparative analysis of several empirical efficiency metrics, including inference time and energy consumption. Experimental results demonstrate that proposed framework is capable of achieving a balance between interpretability, generality and computational efficiency. By leveraging GPU-based batch distillation and parallelized inference, HXRS achieves efficient resource utilization while supporting scalable explainable recommendation.</p>

LLM-guided data distillation for explainable recommender system

期刊论文
发表日期：

2026年3月23日

查看全文

Arunabha Basak,
Bidyut Kr. Patra

摘要

In the era of information overload, recommender systems have become an integral part of our daily lives. These systems can recommend items, products or services to users based on their preferences and past behavior. Explainable recommender systems (XRS) extend this functionality by providing human-understandable justifications for the recommendations. It helps enhance user trust and system transparency. Traditional XRS models primarily relied on small transformer-based language models (SLMs). However, their explanations were often generic and repetitive. Modern XRS frameworks have employed large language models (LLMs) which generate contextually richer explanations. However, relying solely on LLMs makes the model computationally expensive and resource-intensive. To address these limitations, this paper proposes a hybrid XRS framework that combines the strengths of both SLMs and LLMs leveraging data distillation. The framework consists of three phases. In phase one, user and item profiles are constructed by extracting feature information from historical reviews. These profiles become part of the input prompt for data distillation process. In the second phase, raw user reviews are summarized into structured and generalized explanations using an LLM. These serve as the output explanations for the data distillation process. Having built the input–output explanation dataset, lightweight SLMs (GPT-2 and T5-small) are fine-tuned in the last phase. The fine-tuned models are capable of generating rich, interpretable explanations comparable to those of LLMs, with a significant reduction in computational cost. This makes them suitable for scalable and resource-constrained environments. Experiments on three real-world datasets (Yelp, Amazon Fine Foods and Amazon Video Games) demonstrate that our distilled SLMs outperform prior SLM-based frameworks (NRT, PETER, PEPLER) and perform competitively against pure LLM-based frameworks (XRec) across evaluation metrics like BERTScore, BARTScore and BLEURT. In addition, the proposed framework enables low-latency and energy-efficient explanation generation, as evidenced by a comparative analysis of several empirical efficiency metrics, including inference time and energy consumption. Experimental results demonstrate that proposed framework is capable of achieving a balance between interpretability, generality and computational efficiency. By leveraging GPU-based batch distillation and parallelized inference, HXRS achieves efficient resource utilization while supporting scalable explainable recommendation.