PaperlayerView exact revision on arXiv

Source-linked AI summary

Recommender Systems in the Era of Large Language Models (LLMs)

Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, Qing Li

arXiv:2307.02046v6cs.IRcs.AIcs.CL

TL;DR

Existing surveys do not systematically review domain-specific techniques for adapting LLMs to recommender systems. This survey organizes LLM-empowered RecSys research across representation learning, pre-training, fine-tuning, and prompting, providing an in-depth overview and future directions.

  • Problem

    Existing surveys mainly emphasize applications, leaving domain-specific techniques for adapting LLMs to recommender systems insufficiently reviewed.

  • Method

    The survey reviews LLM-empowered recommender systems through user-item representation learning and pre-training, fine-tuning, and prompting paradigms.

  • Results

    The survey provides a comprehensive overview of existing LLM-empowered recommender systems and discusses potential future directions.

  • Takeaways & Limitations

    LLM-based recommender-system research remains at an early stage and requires more systematic and comprehensive studies.

  • Takeaways & Limitations

    LLM-based recommender systems raise privacy risks because they rely on large-scale data and handle sensitive user information.

Abstract

from arXiv · show

With the prosperity of e-commerce and web applications, Recommender Systems (RecSys) have become an important component of our daily life, providing personalized suggestions that cater to user preferences. While Deep Neural Networks (DNNs) have made significant advancements in enhancing recommender systems by modeling user-item interactions and incorporating textual side information, DNN-based methods still face limitations, such as difficulties in understanding users' interests and capturing textual side information, inabilities in generalizing to various recommendation scenarios and reasoning on their predictions, etc. Meanwhile, the emergence of Large Language Models (LLMs), such as ChatGPT and GPT4, has revolutionized the fields of Natural Language Processing (NLP) and Artificial Intelligence (AI), due to their remarkable abilities in fundamental responsibilities of language understanding and generation, as well as impressive generalization and reasoning capabilities. As a result, recent studies have attempted to harness the power of LLMs to enhance recommender systems. Given the rapid evolution of this research direction in recommender systems, there is a pressing need for a systematic overview that summarizes existing LLM-empowered recommender systems, to provide researchers in relevant fields with an in-depth understanding. Therefore, in this paper, we conduct a comprehensive review of LLM-empowered recommender systems from various aspects including Pre-training, Fine-tuning, and Prompting. More specifically, we first introduce representative methods to harness the power of LLMs (as a feature encoder) for learning representations of users and items. Then, we review recent techniques of LLMs for enhancing recommender systems from three paradigms, namely pre-training, fine-tuning, and prompting. Finally, we comprehensively discuss future directions in this emerging field.

1 INTRODUCTION

Recommender systems personalize item suggestions but remain limited in textual understanding and generalization across recommendation scenarios. This survey examines how large language models can address these challenges through pre-training, fine-tuning, and prompting, while systematically reviewing domain-specific adaptation techniques.

  • Motivation: Recommender systems alleviate information overload by providing personalized suggestions tailored to user preferences across domains such as entertainment, e-commerce, and job matching.They model user-item interactions and incorporate textual information, including item descriptions, user profiles, and reviews.
  • Limitations: Existing recommender systems struggle to capture textual knowledge about users and items, resulting in suboptimal prediction performance across recommendation scenarios.The passage attributes this limitation to model scale, data size, and insufficient natural language understanding in models such as CNN, LSTM, and BERT.
  • Limitations: Most existing recommendation methods are task-specific and have inadequate generalization ability to unseen recommendation scenarios.This limitation is presented alongside difficulties in understanding users’ interests and textual side information.
  • LLM Opportunity: LLMs offer potential for recommender systems through their language understanding, generation, generalization, and reasoning abilities, enabled by transformer-based pre-training on vast textual data.The introduction describes LLMs as billion-parameter models trained on diverse sources such as articles, books, websites, and other publicly available written materials.
  • Survey Scope and Novelty: The survey systematically reviews LLM-empowered recommender systems through pre-training, fine-tuning, and prompting, with emphasis on domain-specific techniques for adapting LLMs to recommendations.It aims to provide an in-depth understanding of developing LLM-based methods tailored to recommender systems for future research.

2 RELATED WORK

Related work traces recommender systems from collaborative filtering and content-based methods through deep learning and language-model approaches. It highlights how LLMs’ language capabilities support more personalized recommendations and user-rating prediction.

  • Recommender Systems: Recommender systems address information overload by providing personalized content and services, primarily through collaborative filtering and content-based recommendation.Collaborative filtering predicts future preferences from similar user behavior patterns.
  • Deep Learning for Recommendation: Deep learning methods model user–item interactions and representations, including NeuMF’s nonlinear interactions and GNNs’ graph-structured node representations.DeepCoNN additionally encodes users’ item reviews with parallel CNNs for rating prediction.
  • Language Models for Recommendation: Language models are increasingly used in recommender systems because they comprehend and generate natural language, enabling applications such as news and drug recommendation.BERT4Rec is cited as a sequential recommendation method using bidirectional encoder representations.
  • Pretrained Language Models: Classical pretrained language models include encoder-only BERT, decoder-based GPT, and text-to-text T5, all built on the Transformer architecture.BERT processes bidirectional context, whereas GPT processes word sequences left to right for language generation.
  • Large Language Models: LLMs trained on extensive text with billions of parameters demonstrate strong natural-language understanding and generation, supporting applications across domains including recommender systems.Examples include GPT-3, LaMDA, PaLM, Vicuna, and ChatGPT.

3 DEEP REPRESENTATION LEARNING FOR LLM-BASED RECOMMENDER SYSTEMS

LLM-based recommender systems represent users and items either through discrete IDs or textual side information. ID-based methods enable unified natural-language recommendation formats but lack semantic information, motivating text encoders and vector-quantized representations.

  • ID-based Representation: Users and items are represented as discrete IDs, with LLM-based systems expressing each as a short phrase containing a type prefix and unique ID.These representations model user-item behaviors through embedding vectors of ID representations.
  • ID-based Representation: P5 maps user-item interactions, profiles, item descriptions, and reviews into natural-language sequences using indexed users and items with personalized prompts.P5 uses a pre-trained T5 backbone and transfers multiple recommendation data formats into a unified paradigm.
  • ID-based Representation: Pure ID indexing lacks sufficient semantic information and makes relevance calculation difficult, particularly under severely sparse user-item interactions.The limitation arises because pure ID indexing is naturally discrete.
  • Textual Side Information-enhanced Representation: Textual side information enables language models such as BERT to encode users and items into semantic spaces for grouping similar entities and distinguishing them more finely.Examples include user profiles, reviews, and item titles or descriptions; Unisec learns transferable representations from item descriptions with a lightweight encoder, parametric whitening, and an MoE-enhanced adaptor.
  • Textual Side Information-enhanced Representation: VQ-Rec addresses overemphasis on text features by converting item text into vector-quantized discrete indices and retrieving representations from a code embedding table.The method supplements language-model encoding with discrete item codes for recommendation retrieval.

4 PRE-TRAINING & FINE-TUNING LLMS FOR RECOMMENDER SYSTEMS

This section reviews pre-training and fine-tuning as two paradigms for applying LLMs to recommender systems. Pre-training models general linguistic or behavioral patterns, while fine-tuning adapts pre-trained models to downstream recommendation tasks through full-model, parameter-efficient, and contrastive strategies.

  • Pre-training: Pre-training transfers knowledge from large, diverse unlabeled corpora, enabling LLMs to learn grammar, syntax, semantics, common-sense reasoning, and coherent contextual generation.
  • Pre-training: PTUM uses Masked Behavior Prediction and Next K behavior Prediction to model user behaviors, with NBP predicting future behaviors from interaction histories.NBP captures relevance between past and future behaviors and follows users’ temporal behavior sequences.
  • Fine-tuning: Full-model fine-tuning updates all LLM weights for task specialization, but scaling models beyond 10 billion parameters makes this infeasible on a single consumption-level GPU.Examples include RecLLM for YouTube video recommendation and GIRL for job recommendation; direct fine-tuning may also introduce unintended bias.
  • Pre-training: P5 combines multi-mask modeling with mixed recommendation-task datasets, enabling generalization to diverse and unseen tasks through zero-shot generation.It uses unified user and item indexing in language sequences so masked language modeling can span recommendation tasks.
  • Fine-tuning: Contrastive fine-tuning improves recommendation by learning from intent–anchor–product triples or by combining discriminative matching scores with candidate-text perplexity.
  • Fine-tuning: Parameter-efficient fine-tuning freezes most pre-trained weights and updates adapters or other small trainable components, reducing computational cost while retaining most model capabilities.TallRec aligns LLaMA-7B with recommendation tasks using LoRA and can run on a single RTX 3090.

5 PROMPTING LLMS FOR RECOMMENDER SYS-

Prompting adapts LLMs to downstream recommendation tasks through task-specific text templates that recast tasks as language generation. Recent work explores in-context learning, chain-of-thought, prompt tuning, and instruction tuning, with the latter enhancing zero-shot performance on unseen recommendation tasks.

  • Prompting paradigm: Prompting adapts LLMs to downstream tasks using task-specific text templates and unifies those tasks as language generation.This alignment matches LLM objectives.
  • Prompt design: Recent recommendation research investigates manually designed prompts with In-context Learning (ICL) and Chain-of-thought (CoT).These techniques are explored to facilitate LLM performance across recommendation tasks.
  • Prompt tuning: Prompt tuning adds prompt tokens to LLMs and updates them using task-specific recommendation datasets.It is an additive prompting technique.
  • Instruction tuning: Instruction tuning combines pre-training, fine-tuning, and prompting to tune LLMs across multiple recommendation tasks with instruction-based prompts.This approach enhances zero-shot performance on unseen recommendation tasks.

5.1 Prompting

Prompting adapts frozen LLMs to recommendation tasks through task-specific prompts, progressing from conventional prompt engineering and few-shot prompting to ICL and CoT strategies. These methods can make LLMs recommend directly, bridge LLMs with conventional recommenders, or support autonomous recommendation agents, while performance depends on task structure and reasoning demands.

  • 5.1 Prompting: Prompting keeps LLM parameters frozen and adapts models to downstream tasks through task-specific prompts.Prompting strategies evolved from unifying downstream tasks with language-generation formats toward ICL and CoT for more complex reasoning.
  • 5.1.1 Conventional Prompting: Conventional prompting uses prompt engineering or a few input-output examples to guide pre-trained language models toward desired task outputs.Prompt engineering emulates text encountered during pre-training, whereas few-shot prompting supplies input-output shots.
  • 5.1.1 Conventional Prompting: Conventional prompting has limited recommendation applications because language-generation objectives differ substantially from recommendation tasks.Existing applications mainly address review summarization and relation labeling between items.
  • 5.1.2 In-context Learning (ICL): ICL adapts LLMs through prompts and in-context demonstrations, commonly using few-shot examples or zero-shot task descriptions.Few-shot ICL provides demonstrations with contexts and desired completions, while zero-shot ICL provides only natural-language task descriptions.
  • 5.1.2 In-context Learning (ICL): Few-shot ICL typically outperforms zero-shot ICL, whereas zero-shot ICL removes the need for task-specific recommendation demonstrations and suits some conversational tasks.The comparison arises because few-shot ICL adds in-context demonstrations, while zero-shot ICL does not.
  • 5.1.2 In-context Learning (ICL): ICL can teach LLMs to act as recommenders for top-K recommendation, rating prediction, and explanation generation, or refine candidates from conventional recommendation systems.ChatGPT can receive candidate items from a traditional recommender and refine the final recommendation results.
  • 5.1.3 Chain-of-thought (CoT) Prompting: CoT addresses ICL’s reasoning limitations by eliciting explicit intermediate reasoning steps through zero-shot or few-shot reasoning prompts.Zero-shot CoT generates reasoning steps independently, while few-shot CoT augments demonstrations with manually designed input-CoT-output sequences.
  • 5.1.3 Chain-of-thought (CoT) Prompting: CoT prompting can organize complex recommendations into sub-tasks for autonomous agents and may extend to graph-based reasoning for recommendation tasks.InteRecAgent and RecMind generate plans for using external tools, while graph-based CoT models reasoning steps as nodes connected by reasoning paths.

5.2 Prompt Tuning

Prompt tuning adds and optimizes task-specific prompt tokens, requiring less task-specific knowledge and human effort than manual prompt design while updating only minimal parameters. In RecSys, strategies use either discrete hard prompts or continuous soft prompts, with soft prompts also incorporating learned representations.

  • 5.2 Prompt Tuning: Prompt tuning adds new prompt tokens to LLMs and optimizes them on task-specific datasets with minimal updates to the tunable prompt and input layer.Compared with manually generating task-specific natural-language prompts, it requires less task-specific knowledge and human effort.
  • 5.2 Prompt Tuning: Prompt tuning strategies for RecSys are categorized into hard prompt tuning and soft prompt tuning according to whether prompts are discrete or continuous.Discrete prompts are also called hard prompts, while continuous prompts are called soft prompts.
  • 5.2.1 Hard Prompt Tuning: Hard prompt tuning generates and updates discrete natural-language prompt templates for specific downstream recommendation tasks.In-context learning can be considered a subclass of hard prompt tuning.
  • 5.2.1 Hard Prompt Tuning: Hard prompt tuning faces discrete optimization challenges, requiring laborious trial and error to discover suitable prompts across a vast vocabulary space.The optimization process seeks prompts suitable for specific recommendation tasks.
  • 5.2.2 Soft Prompt Tuning: Soft prompt tuning uses continuous vectors such as text embeddings, optimizes them with task-specific recommendation losses, and updates only soft prompts and minimal input-layer parameters.Soft prompt tokens are commonly concatenated with original input tokens at the tokenizer input layer.
  • 5.2.2 Soft Prompt Tuning: Existing works improve soft prompts by encoding task-specific representations, including contrastive-learning-based user representations and mutual information for cross-domain recommendations.Wu et al. encode user representations into prompt tokens, while Wang et al. and Guo et al. encode cross-domain mutual information.

5.3 Instruction Tuning

Instruction tuning addresses LLMs’ weaker zero-shot performance by combining prompting and fine-tuning across multiple task-specific prompts. It consists of instruction generation followed by model tuning for downstream recommendation tasks.

  • Instruction Tuning: Instruction tuning combines prompting with fine-tuning LLMs over multiple task-specific prompts to address limited zero-shot ability.The supplied passage contrasts prompting’s few-shot performance on unseen tasks with its poorer zero-shot ability.
  • Instruction (Prompt) Generation Stage: Instruction generation creates natural-language instruction prompts pairing task-oriented inputs with desired targets from task-specific datasets.For recommendation, templates can include user preferences, intentions, task forms, task descriptions, inputs, and outputs.
  • Model Tuning Stage: Model tuning fine-tunes LLMs over multiple instructions for downstream recommendation tasks using full-model or parameter-efficient tuning.Existing recommender-system works are categorized by these two fine-tuning manners; LoRA is given as an example of parameter-efficient tuning for LLaMA.

6 FUTURE DIRECTIONS · 6.1 Hallucination Mitigation · 6.2 Trustworthy Large Language Models for Recommender Systems

Future work for LLM-enhanced recommender systems should address hallucination and the trustworthiness challenges arising from unreliable, biased, opaque, privacy-sensitive, and adversarially vulnerable models. Promising directions include factual knowledge supplementation, output verification, fairness research, explainability, and privacy-preserving federated learning.

  • 6 FUTURE DIRECTIONS: LLM adaptation to recommender systems remains at an early stage, leaving substantial challenges and opportunities for future research.
  • 6.1 Hallucination Mitigation: Hallucinations produce plausible-sounding but factually incorrect or input-unreferable outputs, motivating factual knowledge supplementation and output verification.Factual knowledge graphs can support training and inference, while outputs can be scrutinized for accuracy and factuality.
  • 6.2 Trustworthy Large Language Models for Recommender Systems: LLM-based recommender systems may threaten users and society through unreliable decisions, unequal treatment, limited transparency and explainability, and privacy issues.
  • 6.2.1 Safety&Robustness: LLMs are highly vulnerable to adversarial perturbations that compromise safety and robustness, including malicious noisy inputs intended to generate unlawful profits or manipulate product markets.
  • 6.2.2 Non-discrimination&Fairness: LLMs can perpetuate biases and stereotypes from human data, causing discriminatory recommendations and unfair treatment of certain user groups.Existing fairness studies include FaiRLLM and UP5, while Hou et al. use prompts to improve item-side fairness; research remains preliminary.
  • 6.2.3 Explainability: Closed LLM architectures and parameters make recommender systems black boxes, complicating users’ understanding of why particular outputs or recommendations are produced.
  • 6.2.4 Privacy: Privacy risks arise because LLM-based recommender systems rely on large, diverse datasets that may contain sensitive user information used for training and personalization.Federated-learning frameworks offer opportunities for decentralized, localized learning that can support more privacy-preserving recommender systems.
  • 6.2.4 Privacy: Federated learning can enable localized learning over distributed data sources, supporting decentralized recommender systems with stronger data privacy.

6.3 Vertical Domain-Specific LLMs for Recommender Systems

Vertical domain-specific LLMs are trained and optimized for particular domains or industries, making them more focused on domain knowledge and skills than general LLMs in recommender systems. This focus gives domain-specific LLM-empowered recommender systems greater domain expertise and practicality.

  • 6.3 Vertical Domain-Specific LLMs for Recommender Systems: Vertical domain-specific LLMs are trained and optimized for specific domains or industries, including healthcare and finance.The passage contrasts these models with general LLMs such as ChatGPT.
  • 6.3 Vertical Domain-Specific LLMs for Recommender Systems: Compared with general LLMs for recommender systems, vertical domain-specific models focus more strongly on the knowledge and skills of a particular domain.Their specialization targets domain-specific expertise rather than universal applicability.
  • 6.3 Vertical Domain-Specific LLMs for Recommender Systems: This domain focus gives LLM-empowered recommender systems greater domain expertise and practicality.The passage explicitly identifies domain expertise and practicality as advantages over general LLM-based recommender systems.

6.4 Users&Items Indexing

LLMs may struggle to capture user–item interaction information from long texts, motivating advanced indexing methods that preserve the collaborative knowledge in users’ and items’ unique identities.

  • Challenge: LLMs may not effectively capture user–item interaction information when interactions are represented in long texts.The passage identifies long-text processing as a challenge for recommender systems.
  • Indexing approach: Advanced methods for indexing users and items are needed beyond merely representing them in text formats.User–item interactions such as clicks, likes, and subscriptions use unique discrete identities that contain rich collaborative knowledge.

6.5 Fine-tuning Efficiency · 6.6 Data Augmentation

Fine-tuning adapts pre-trained LLMs to specific recommendation tasks while retaining their general language understanding, but can be computationally expensive for very large models and datasets. Data augmentation addresses the resource intensity and long-term sustainability concerns of conventional real-data-driven recommender-system research, whose performance and versatility depend on input-data quality and variety.

  • 6.5 Fine-tuning Efficiency: Fine-tuning adapts a pre-trained LLM to a specific recommendation task or domain.Examples include recommending movies or books.
  • 6.5 Fine-tuning Efficiency: This adaptation preserves the model’s general language understanding capabilities learned during pre-training.Fine-tuning simultaneously specializes the model’s knowledge for the target task.
  • 6.5 Fine-tuning Efficiency: Fine-tuning can be computationally expensive in recommender systems, particularly with very large models and large datasets.The passage identifies computational cost as a limitation of applying LLMs to recommendation.
  • 6.6 Data Augmentation: Conventional recommender-system studies commonly rely on real data collected through digital-platform user interactions or annotator recruitment.These are the two data-collection routes identified in the passage.
  • 6.6 Data Augmentation: Real-data-driven approaches are resource-intensive and may not be sustainable over the long term.The passage presents these concerns as shortcomings motivating data augmentation.
  • 6.6 Data Augmentation: Input-data quality and variety directly influence recommender-model performance and versatility.This relationship motivates efforts to overcome limitations of conventional real-data collection.

7 CONCLUSION

LLMs’ language understanding, generation, generalization, reasoning, and prompt-adaptation abilities have motivated efforts to improve recommender systems. This survey addresses the need for a systematic overview by organizing LLM-empowered recommender systems around pre-training, fine-tuning, and prompting.

  • LLMs have achieved success across applications through language understanding and generation, generalization and reasoning skills, and prompt adaptation to new tasks and domains.
  • Researchers increasingly use LLMs to improve recommender systems and provide high-quality, personalized suggestion services.
  • The survey provides a comprehensive overview of LLM-empowered recommender systems through pre-training, fine-tuning, and prompting paradigms.
Loading 2307.02046v6…