PaperlayerView exact revision on arXiv

Source-linked AI summary

Cold-Start Recommendation towards the Era of Large Language Models (LLMs): A Comprehensive Survey and Roadmap

Weizhi Zhang, Yuanchen Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, Feiran Huang, Sheng Zhou, Jiajun Bu, Allen Lin, James Caverlee, Fakhri Karray, Irwin King, Philip S. Yu

arXiv:2501.01945v2cs.IRcs.AI

TL;DR

Cold-start recommendation lacks a comprehensive account of how to model new or interaction-limited users and items across diverse knowledge sources. This survey reviews the field from content features, graph relations, and domain information toward LLM-based world knowledge, while defining cold-start tasks and outlining future directions. It concludes with an organized roadmap and community resource, while identifying deployment, privacy, and evaluation constraints.

  • Problem

    Cold-start recommendation remains difficult because new users and items have little or no information for modeling user-item similarity, while existing surveys cover only partial knowledge scopes or limited cold-start aspects.

  • Method

    The paper conducts a comprehensive survey that defines cold-start problems, categorizes methods across four knowledge scopes, reviews representative work, and proposes a roadmap.

  • Results

    The survey provides a comprehensive review of cold-start recommendation research, formalizes different research questions, and organizes related resources for the research and industrial community.

  • Takeaways & Limitations

    The roadmap frames cold-start recommendation as progressing from content, graph, and domain information toward world knowledge from large language models.

  • Takeaways & Limitations

    Current approaches primarily target offline settings and face high latency, computational overhead, and resource costs in large-scale real-time deployment.

Abstract

from arXiv · show

Cold-start problem is one of the long-standing challenges in recommender systems, focusing on accurately modeling new or interaction-limited users or items to provide better recommendations. Due to the diversification of internet platforms and the exponential growth of users and items, the importance of cold-start recommendation (CSR) is becoming increasingly evident. At the same time, large language models (LLMs) have achieved tremendous success and possess strong capabilities in modeling user and item information, providing new potential for cold-start recommendations. However, the research community on CSR still lacks a comprehensive review and reflection in this field. Based on this, in this paper, we stand in the context of the era of large language models and provide a comprehensive review and discussion on the roadmap, related literature, and future directions of CSR. Specifically, we have conducted an exploration of the development path of how existing CSR utilizes information, from content features, graph relations, and domain information, to the world knowledge possessed by large language models, aiming to provide new insights for both the research and industrial communities on CSR. Related resources of cold-start recommendations are collected and continuously updated for the community in https://github.com/YuanchenBei/Awesome-Cold-Start-Recommendation.

1 INTRODUCTION

Cold-start recommendation addresses sparse or absent interaction data for new users and items, and this survey organizes the field by expanding knowledge sources from content and graphs to domains and LLM-derived world knowledge. It reviews prior work, defines a broader task landscape, and proposes a roadmap for future research.

  • Motivation: Cold-start recommendation is increasingly important because new users, items, and platforms often have limited or no historical interaction data.The survey links this challenge to fairness, diversity, and the effective sustainability of recommender systems.
  • Knowledge scopes: Existing cold-start methods progress from content features such as titles, genres, and profiles toward graph relations and cross-domain information transfer.Graph-based approaches use interaction, knowledge, and social networks, while cross-domain methods exploit richer data from other domains.
  • Survey scope: The survey categorizes cold-start recommendation research into four knowledge scopes: Content Features, Graph Relations, Domain Information, and LLM-based World Knowledge.This taxonomy is intended to organize how external knowledge addresses data and interaction scarcity.
  • Related work: No prior survey fully covers all four knowledge scopes while addressing multiple cold-start issues, motivating this review of 220 papers through December 2024.The paper also defines nine distinct cold-start issues and compares its coverage with earlier surveys.
  • Method and contributions: The authors use a semi-systematic search across academic databases and major conferences, then analyze motivations, data requirements, and technical approaches.The contribution includes a timeline and statistical overview of publications alongside the taxonomy and roadmap.

2 PRELIMINARIES

Recommender systems infer user preferences from user-item interactions, but cold-start settings make this difficult because new users and items provide little or no information. The survey formalizes cold-start tasks and distinguishes nine scenarios across four categories.

  • Recommender systems: Recommender systems predict user preferences or interaction likelihoods, primarily from historical user-item interactions.They support item discovery across e-commerce, social media, and content platforms.
  • Cold-start problem: Cold-start recommendation concerns accurate recommendations for new users or items whose sparse information makes user-item similarity difficult to model.The paper presents this as a long-standing challenge for recommender-system research.
  • Survey framework: The survey reviews cold-start methods across knowledge scopes ranging from content features, graph relations, and domain information to world knowledge.It combines formal problem definitions with classification and discussion of existing models.
  • Task definition: Nine cold-start recommendation tasks are grouped into long-tail, normal cold-start, strict cold-start, and system cold-start categories.These categories differ according to how users or items are observed and how training, tuning, and testing sets are formed.

3 CONTENT FEATURES

Content features use descriptive information about users and items to represent cold entities when interaction histories are scarce or absent. The survey distinguishes data-incomplete learning for strict cold-start settings from data-efficient learning for normal cold-start settings with limited interactions.

  • Content features include user profiles, user reviews, item names, and descriptions that characterize user or item attributes.
  • Content-feature methods address either strict cold-start scenarios without prior interactions or normal cold-start scenarios with limited interaction data.

3.1 Data-Incomplete Learning

Data-incomplete learning relies on content information to represent strictly cold users or items and align them with warm representations learned from historical interactions. The survey groups these methods into co-training, knowledge alignment, exploration, and feature-similarity approaches.

  • Data-incomplete learning uses only content information to represent cold users or items and unify them with warm representations learned from historical interactions.
  • Robust Co-Training: Robust co-training jointly uses behavior-based warm representations and content-based cold representations, applying robust generalization strategies to support cold situations.
  • Knowledge Alignment: Knowledge alignment addresses semantic discrepancies by aligning content-based cold representations with behavior-based warm representations through contrastive learning, distillation, or adversarial training.
  • Knowledge Alignment: Contrastive learning bridges cold content representations and warm behavior representations by bringing related instances closer while separating dissimilar ones.
  • Cold Exploration: Cold exploration uses recommender feedback to rapidly adjust representations of cold users or items, including reinforcement-learning strategies for preference discovery.
  • Feature Similarity Measurement: Feature-similarity methods model cold users or items through content similarity, while multi-feature fusion and hashing combine or unify representations for measurement.

3.2 Data-Efficient Learning

Data-efficient learning targets normal cold-start settings where limited interactions are available, with meta-learning pretraining diverse users and rapidly adapting models to new users or items. The survey further covers task-aware adaptation, meta-embeddings, and sequential meta-learning.

  • Normal cold-start recommendation improves learning from limited user-item interactions through meta-learning and related adaptation strategies.
  • Meta-learning pretrains models on diverse users’ historical interactions and rapidly adapts them to new cold-start users or items.
  • Pretraining: Pretraining methods simulate few-shot test scenarios, using support-set updates for decision layers and query-set updates for the broader preference estimator.
  • Adaptation: Adaptation methods personalize cold-start updates with feature-specific and task-specific memories while addressing risks of memorizing query interactions.
  • Task Similarity: Task-aware meta-learning considers task differences and relevance because treating users as isolated tasks limits recognition of user contributions and task relationships.
  • Sequential Meta-Learning: Sequential meta-learning incorporates the time order of limited interactions to capture dynamic preferences and match cold-start items with potential users.

4 GRAPH RELATIONS

Graph-relation methods compensate for sparse or absent interactions by enhancing interaction graphs, extending them with richer relations, or improving graph aggregation. These strategies provide cold nodes with additional structural or associated information.

  • Graph neural recommender models update node embeddings by aggregating neighborhood information through message passing.
  • Interaction Graph Enhancement: Interaction graph enhancement increases graph information for cold nodes through supplementary relations or homophily-based network relations.
  • Interaction Graph Enhancement: Supplementary graph relations add edges for cold instances, allowing external information from other nodes to be aggregated when generated edges are sufficiently suitable.
  • Interaction Graph Enhancement: Homophily-based methods exploit explicit or implicit associations among users and items under the assumption that neighboring nodes have similar behaviors or labels.
  • Graph Relation Extension: Graph relation extension adds heterogeneous, attributed, or knowledge-graph relations because the original interaction graph lacks sufficient information for cold nodes.
  • Graph Relation Extension: Heterogeneous graphs expand node and edge types to provide more complex relationships, while knowledge graphs supply semantic entity relations and auxiliary information for cold-instance learning.

4.3 Graph Aggregator Improvement

Graph aggregator improvement enhances cold-start recommendation by extracting more usable information from limited structural data. It either broadens the model’s receptive field or refines aggregation to prioritize critical information for cold nodes.

  • Graph aggregator improvement is a model-centric approach for extracting more usable information from limited structural data in cold-start recommendation.
  • Global or long-range context: Global-context methods extend beyond local neighborhoods so cold instances can use long-range correlated nodes despite sparse direct neighborhoods.MeGNN combines global and local neighborhood transformation learning, while MPT uses a Transformer encoder within a GNN to capture long-range dependencies.
  • The two categories differ between expanding the model’s scope to distant context and improving aggregation within limited local information.
  • Aggregator refinement: Aggregator-refinement methods improve the information captured from limited cold-instance interactions.A self-attention meta-aggregator addresses cold-start neighbors, while A-GAR adaptively explores higher-order information.

5 DOMAIN INFORMATION

Domain information is used as a source of knowledge for cold-start recommendation, especially when target-domain interactions are sparse. Cross-domain methods leverage richer source-domain information through relationships among domains.

  • Cross-domain information is presented as a distinct category of knowledge used to address cold-start recommendation.

5.1 Domain Knowledge Transfer

Domain knowledge transfer moves information from warm source domains to cold target domains through mappings, heterogeneous connections, and training or tuning techniques. These approaches range from general representation alignment to personalized, graph-based, and meta-learning-based transfer.

  • Domain transfer methods transfer knowledge from warm source domains to cold target domains through embedding mapping, graph connections, or learning processes.
  • Embedding and feature mapping: Embedding and feature-transfer methods align representation spaces across domains, commonly using nonlinear transformations.General mapping often uses MLPs, while MAFT combines MLPs with attention for auxiliary feature spaces.
  • Personalized mapping: Personalized mapping generates user-specific bridging functions rather than applying one universal mapping.PTUPCDR uses a meta-network to generate personalized parameters from source-domain user embeddings, while VRCDR uses interaction-pattern characteristic vectors.
  • Heterogeneous connections: Heterogeneous approaches construct graph-based connections to explicitly model and transfer knowledge across domains.Knowledge-graph methods use meta-path aggregation or natural item relationships, while hybrid graphs connect users or items across domains.
  • Training and tuning: Training and tuning techniques implicitly pass information between source and target domains through joint training, transfer models, or efficient adaptation.Meta-learning approaches pre-train on source and target domains before tuning a task-specific meta-network for cold-start generalization.

5.2 Domain Distribution Alignment

Domain distribution alignment reduces differences between source and target domains so shared features, behaviors, or auxiliary information can support cold-start recommendation. Methods align collaborative-filtering patterns, latent representations, or auxiliary embeddings using contrastive, adversarial, and Stein-path techniques.

  • Domain alignment reduces distributional differences between source and target domains to enable knowledge sharing in cold-start recommendation.
  • Collaborative-filtering alignment: Collaborative-filtering alignment uses shared interaction patterns and user behaviors to improve transfer to cold users in target domains.
  • Contrastive and latent alignment: Contrastive alignment aligns users or domain samples, while latent-dimension alignment matches representations in encoder-decoder spaces.DAUC combines contrastive and adversarial alignment losses, and LACDR aligns overlapping user representations in a low-dimensional space.
  • Auxiliary-feature alignment: Auxiliary-feature alignment makes cross-domain alignment more approachable than directly matching latent embedding distributions.
  • Stein-path alignment: Stein-path methods iteratively move target embeddings toward source distributions to form semantic bridges between related items.DisAlign uses probabilistic Stein paths, such as aligning a book’s auxiliary embedding with its movie adaptation.
  • Robust alignment: Contrastive augmentation and intra- and inter-domain learning are used to improve robustness and address data sparsity during domain alignment.CCDR aligns users, taxonomies, and neighbors while creating augmented sub-graphs within domains.

5.3 Domain-Invariant Representation Learning

Domain-invariant representation learning assumes shared features can transfer across domains while preserving domain-specific traits. The survey groups methods by how they disentangle or fuse these shared and domain-specific representations.

  • Core idea: Domain-invariant learning separates shared features capturing transferable preferences or item characteristics from domain-specific features.The shared and domain-specific components are optimized to retain complementary information across domains.
  • Adversarial Learning: Adversarial methods train discriminators and generators to produce representations that obscure domain identity while isolating domain-invariant features.Diff-MSR extends this paradigm with diffusion models to generate domain-specific embeddings.
  • Attention Mechanism: Attention mechanisms disentangle representations by modeling feature-field relationships and aligning domain-shared features across heterogeneous graphs.Some approaches combine attention with cross-view contrastive learning.
  • Multi-View Learning: Multi-view learning integrates complementary semantic and structural perspectives to construct shared representations for cross-domain recommendation.Examples combine descriptive content semantics with high-order knowledge-graph connectivity.
  • Swapping Learning: Swapping learning exchanges domain-specific information to refine shared features and filter irrelevant signals.Reported approaches swap domain-specific or aspect-level preferences across domains.
  • Semantic Learning: Semantic learning maps auxiliary information such as reviews, browsing histories, and item descriptions into a shared semantic space for knowledge transfer.These approaches treat semantic features as domain-invariant information.

6 WORLD KNOWLEDGE FROM LARGE LANGUAGE MODELS

The survey traces how LLM world knowledge is used for cold-start recommendation, covering LLMs as recommenders and as knowledge enhancers. It organizes methods around prompting, retrieval, tuning, and representation enhancement for zero-shot, few-shot, and sparse-interaction settings.

  • Overview: LLMs learn broad world knowledge from large text corpora and can generate text, answer questions, translate, and conduct complex conversations.Their textual and multimodal processing capabilities motivate their use in recommendation.
  • Overview: The survey categorizes LLM-based CSR methods as LLMs acting as recommender systems or as knowledge enhancers.The two categories organize how LLM capabilities enter the recommendation pipeline.
  • LLM as the Recommender System: Prompting methods frame recommendation as natural-language processing using direct prompts, in-context examples, multi-step prompts, and retrieved information.These strategies target zero-shot and few-shot recommendation without necessarily retraining the model.
  • LLM as the Recommender System: Retrieval-augmented recommendation retrieves item characteristics, candidate items, or interaction histories before prompting an LLM-based recommender.The retrieved fine-grained information is intended to strengthen modeling of cold and zero-shot instances.
  • LLM as the Recommender System: Instruction tuning and fine-tuning adapt pretrained LLMs to recommendation by learning from textualized histories, metadata, recommendation losses, or additional trainable parameters.These methods aim to encode collaborative and semantic information more explicitly.
  • LLM as the Knowledge Enhancer: LLMs enhance cold-instance representations through modality-enhanced and domain-enhanced representations, extending information beyond sparse ID embeddings.Auxiliary linguistic or multimodal encoders help align richer content with recommendation representations, while domain enhancement targets cross-domain transfer.

7 CHALLENGES AND FUTURE OPPORTUNITIES

The survey identifies future opportunities across algorithm development, deployment, and benchmarking for cold-start recommendation. It highlights missing modalities, foundation-model adaptation, industrial efficiency, privacy, and inconsistent evaluation as key challenges.

  • 7.1 Multi-Modal Cold-Start Recommendation: Multi-modal information can provide richer user and item representations, but real-world systems often lack modalities and datasets provide limited multimodal coverage.Existing models may also mismatch content understanding with personalization and overlook users’ differing modality sensitivities.
  • 7.2 Recommendation Foundation Models: Foundation models could support multiple CSR subtasks and dynamically adjust recommendations across diverse domains.The survey points to adapting foundation models beyond individual recommendation tasks.
  • 7.3 Efficiency in Cold-Start Recommendations: Current CSR approaches are mainly evaluated offline and face high latency and resource costs in large-scale real-world deployment.The survey calls for lightweight, scalable, hybrid, and real-time learning strategies.
  • 7.4 Data Privacy in Cold-Start Recommendations: Cold-start methods rely on user information beyond interactions, intensifying privacy challenges involving side information and cross-domain data.The survey identifies privacy-preserving technologies such as differential privacy as promising directions.
  • 7.5 Benchmark and Unified Evaluation: CSR evaluations remain diverse and inconsistent, motivating unified and fair benchmarks across strict, non-strict, and long-tail cold-start settings.The survey also notes inconsistent dataset choices, unclear cold-entity definitions, and the need for open evaluation frameworks.

8 CONCLUSION

The paper provides a comprehensive review and roadmap for cold-start recommendation, tracing progress from conventional information sources toward world knowledge from large language models. It also identifies challenges, future directions, and community resources.

  • The paper reviews cold-start recommendation through a roadmap spanning content features, graph relations, domain information, and large-language-model world knowledge.
  • It formally defines cold-start research questions and systematically organizes representative works with overarching insights.
  • The paper reconsiders cold-start challenges, summarizes future directions, and organizes related resources for the research and industrial community.
Loading 2501.01945v2…