Comparative analysis of methodologies and approaches in recommender systems utilizing large language models

Artificial Intelligence Review (2025) 58:207 https://doi.org/10.1007/s10462-025-11189-8 Comparative analysis of methodologies and approaches in recommender systems utilizing large language models Salma S. Elmoghazy1 · Marwa A. Shouman2 · Hamdy K. Elminir1 · Gamal Eldin I. Selim2 Accepted: 6 March 2025 © The Author(s) 2025 Abstract Recommendation systems are indispensable technologies nowadays, as they enable analysis of the huge amount of information available on the internet, helping consumers to make decisions effectively. Ongoing efforts are essential to further develop and align them with the evolving demands of the modern era. In the last few years, large language models (LLMs) have made a huge leap in natural language processing. This advancement has directed researchers’ efforts towards employing these models in various fields, including recommender systems, to leverage the vast amount of data they were trained on. This paper presents a comparative study of a set of recent methodologies that adapt LLMs to recommendations. Throughout the discussed research work, we come up with the insight that LLMs offer significant benefits due to the amount of knowledge they possess and their powerful ability to represent textual data effectively, making them useful in common recommendation issues like cold-start. Also, the variety of fine-tuning and in-context learning techniques enables adaptation of LLMs to a wide range of recommendation tasks. We discussed issues addressed in the reviewed research work and the solutions proposed to enhance recommendation systems. To provide a clearer understanding, we propose taxonomies to categorize the reviewed work based on underlying techniques, involving the role of LLMs in recommendations, learning paradigms, and system structures. We explore datasets, recommendation- and language-related metrics commonly used in this domain. Finally, we analyzed findings in related work, highlighting possible strengths and limitations of using LLMs in recommender systems. Keywords Recommender systems · Large language models (LLMs) · Natural language processing (NLP) · User modeling · Collaborative filtering (CF) · Parameter-efficient fine-tuning (PEFT) Marwa A. Shouman, Hamdy K. Elminir, and Gamal Eldin I. Selim have contributed equally to this work. Extended author information available on the last page of the article 13 207 Page 2 of 36 S. S. Elmoghazy et al. 1 Introduction In today’s digital world, where information grows exponentially, recommender systems (RSs) are essential to assist users in finding content, services, and products that align with their preferences. From movie suggestions on Netflix to personalized news feeds on social media, recommender systems have reshaped users’ online experiences. Collaborative filtering is one of the earliest and most popular techniques of recommendation. It assumes that similar users tend to have similar interests (Su and Khoshgoftaar 2009). As an enhancement of collaborative filtering techniques, matrix factorization methods gained a lot of attention during the Netflix Prize competition, allowing efficient scalability for recommender systems (Koren et al. 2009). Another popular type of recommender systems is content-based filtering methods that depend on the similarity between the users past consumed items and the candidate items (Mooney and Roy 2000). Hybrid recommender systems were introduced to leverage the capabilities of both collaborative filtering and content-based filtering, through various implementation methods, to overcome the shortcomings of both approaches (B.Thorat et al. 2015). In conjunction with the remarkable developments in the field of deep learning, many sophisticated neural network-based recommendation techniques have evolved (Zhang et al. 2019; He et al. 2017; Fan et al. 2022), resulting in a huge advancement in recommendation systems. Following the advancements in deep learning-based recommender systems, the integration of natural language processing (NLP) techniques, such as recurrent neural networks-based (Hidasi et al. 2016) and attention-based (Kang and McAuley 2018) techniques, has emerged, providing better modelling for user’s historical sequences of behaviors. However, like any technology, recommender systems face various challenges such as data sparsity, cold-start, scalability, and lack of explainability problems (B.Thorat et al. 2015). Very recent research has started to explore the potential of the integration of large language models (LLMs) like OpenAI’s GPT (Radford et al. 2018) series, T5 (Raffel et al. 2020), BERT (Devlin et al. 2019), BART (Lewis et al. 2020) etc., to further mitigate these limitations. LLMs, trained on massive textual datasets, possess remarkable zero-shot and few-shot capabilities, which enable them to generate responses to tasks they were not explicitly trained on (Brown et al. 2020). This gives the potential to employ LLMs in recommendation systems to solve data sparsity and cold-start problems (Sanner et al. 2023). LLMs have proven their impressive capabilities in adapting to downstream tasks. The same principle applies to recommendation tasks by tuning LLMs using recommendation data (Bao et al. 2023). This flexibility opens the way towards developing unified or multi-tasking recommendation frameworks like P5 (Geng et al. 2022) and M6-Rec (Cui et al. 2022). The growing focus on efficient fine-tuning techniques has the potential to enable LLMs to scale up to large datasets of users and items (Han et al. 2024; Katlariwala and Gupta 2024). Unlike traditional recommender systems, which rely on implicit signals like clicks and interactions, which can be misleading, LLMs excel at capturing semantic information from textual data. This enables the use of descriptions, reviews, and other textual information associated with items and users, helping to obtain better representations for both users and items necessary for improving recommendation performance (Qiu et al. 2021). Additionally, LLM-based recommender systems can be game-changers when it comes to personalized recommendations (Katlariwala and Gupta 2024; Geng et al. 2022). The unique reasoning abilities of LLMs enable them to capture the implied user intent based on the context, which 13 Comparative analysis of methodologies and approaches in… Page 3 of 36 207 traditional CF models often overlook. This results in recommendations that are not only personalized but also contextually relevant, enhancing user satisfaction. Another advantage of LLMs is that they can generate textual explanations and summaries, which can be leveraged to improve the explainability of recommendations, making the recommender system more user-trusted (Li et al. 2023). In the later sections of this survey, we will further explore various research projects that leverage LLMs in recommendation tasks. The goal of this survey is to provide a comparative analysis of selected approaches that use LLMs in recommend (...truncated)