冷启动推荐：系统性综述

原论文链接：

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

CONTENT FEATURES

• 数据不完整学习（Data-Incomplete Learning）
  • 稳健协同训练（Robust Co-Training ）
    • 稳健泛化（Robust generalization）: 这类模型通过稳健的泛化策略，同时优化基于行为的表征和基于内容的表征
      • 2017.Dropoutnet: Addressing cold start in recommender systems
      • 2020.Recommendation for new users and new items via randomized training and mixture-of-experts transformation.
      • 2020.How to learn item representation for cold-start multimedia recommendation?.
      • 2022.Transform cold-start users into warm via fused behaviors in large-scale recommendation.
      • 2024.Temporally and distributionally robust optimization for cold-start recommendation.
    • 自编码（Autoencoders）: 自编码技术采用编-解码架构，通过变分或去噪策略训练的编码器与负责信息重建的解码器联合训练，从而实现对冷启动和热实例的统一表征。
      • 2019.From zero-shot learning to cold-start recommendation
      • 2021.Zero shot on the cold-start problem: Modelagnostic interest learning for recommender systems.
      • 2022.Improving item cold-start recommendation via model-agnostic conditional variational autoencoder.
      • 2023.Gorec: a generative cold-start recommendation framework.
      • 2024.Collaborative Filtering in Latent Space: A Bayesian Approach for Cold-Start Music Recommendation.
  • 知识对齐（Knowledge Alignment） : 知识对齐的核心在于对齐器（aligner），其目标是通过对齐策略将内容特征生成的冷表征与行为数据产生的热表征进行校准。通过这种方式将热表征中蕴含的有意义行为信息迁移至冷表征
    • 对比学习（Contrastive learning ）
      • CLCRec ： 2021. Contrastive learning for cold-start recommendation.
      • CCFCRec ： 2023. Contrastive collaborative filtering for cold-start item recommendation.
    • 知识蒸馏（knowledge distillation） : 与传统的知识蒸馏（KD）应用不同，在冷启动推荐场景中，该方法通常用于从基于行为数据的热表征中提取知识，并将其迁移至基于内容的冷表征。
      • ALDI ： 2023. Aligning distillation for cold-start item recommendation.
      • DTKD ： 2023. Dual-Teacher Knowledge Distillation for Strict Cold-Start Recommendation.
    • 生成式对抗网络（generative adversarial networks ）: 生成对抗网络（GANs）是一类通过生成器与判别器双网络对抗训练的方法，使生成器能持续输出接近真实的数据。在冷启动推荐中，GANs 通常用于使基于内容特征生成的冷表征更接近推荐系统基于行为的的热表征。
      • GAR ： 2022. Generative adversarial framework for cold-start item recommendation.
      • GF2 ： 2022. Revisiting Cold-Start Problem in CTR Prediction: Augmenting Embedding via GAN.
      • GAZRec ： 2022. Generative adversarial zero-shot learning for cold-start news recommendation
  • 冷试探（Cold Exploration） : 在缺乏足够交互数据来有效建模冷启动用户或物品的情况下，采用基于强化学习等"试探性"方法是一种自然而直观的解决方案。
    • Reinforcement learning:
      • RL-LTV ：2021. Reinforcement learning to optimize lifetime value in cold-start recommendation.
      • MetaCRS： 2023. Meta policy learning for cold-start conversational recommendation.
      • WSCB： 2023. User cold-start problem in multi-armed bandits
  • 特征相似性度量（Feature Similarity Measurement ）：为克服缺乏行为数据时的难题，一种替代方案是将重点转向用户和物品的内容特征的表征与建模。具体而言，这类特征相似性度量方法从内容特征相似性的角度来学习和评估用户/物品兴趣。通过这种方式，模型能够规避热表征（来自行为数据）与冷表征（来自内容数据）之间的信息差异。
    • 多特征融合（Multi-feature fusion）：该方法通过综合利用多重特征，缓解数据不完整问题，从而为冷启动实例的度量提供更丰富的信息支持
      • 2011.Collaborative topic modeling for recommending scientific articles.
      • 2016.Solving cold-start problem in large-scale recommendation engines: A deep learning approach.
      • 2022.SMINet: State-aware multi-aspect interests representation network for cold-start users recommendation.
      • 2023.Cross-modal content inference and feature enrichment for cold-start recommendation.
      • 2023.Automatic Fusion Network for Cold-start CVR Prediction with Explicit Multi-Level Representation.
    • Hashing ： 哈希算法在计算机视觉和多媒体领域已被广泛用于相似内容检索，以平衡检索效果与效率，在冷启动推荐场景中，哈希技术可将热表征和冷表征映射到统一的二进制哈希码空间，从而实现相似性度量
      • 2016.Deep supervised hashing for fast image retrieval.
      • 2019.Deep incremental hashing network for efficient image retrieval
      • 2023.Multi-modal hashing for efficient multimedia retrieval: A survey
• 数据有效学习（Data-Efficient Learning）
  • 元学习（meta-learning optimization）: 推荐系统中元学习的核心在于首先通过多样的用户历史交互数据对模型进行预训练，然后利用有限的新增的交互数据快速适应冷启动用户或物品。
    • 预训练（Pretraining）
      • 2019.Melu: Meta-learned user preference estimator for cold-start recommendation.
      • 2019.Meta-learning for user cold-start recommendation
      • 2021.FORM: follow the online regularized meta-leader for cold-start recommendation. I
      • 2022.PNMTA: A pretrained network modulation and task adaptation approach for user cold-start recommendation.
    • 适配（Adaptation ）
      • 2020.Mamo: Memory-augmented meta-optimization for cold-start recommendation.
      • 2021.CMML: Contextual modulation meta learning for cold-start recommendation.
      • 2022.PNMTA: A pretrained network modulation and task adaptation approach for user cold-start recommendation.
      • 2023.A Preference Learning Decoupling Framework for User Cold-Start Recommendation.
  • 元任务（meta-task utilization）：元学习中优化中，多项研究工作特别强调了任务相似性与差异性的关键作用
    • 任务差异性（Task difference ）： 忽视任务差异性可能导致预测偏差，因为具有高度不确定性或高难度的用户可能会对预测结果产生不成比例的影响
      • 2023.Modeling Preference as Weighted Distribution over Functions for User Cold-start Recommendation.
      • 2023.Task-difficulty-aware meta-learning with adaptive update strategies for user cold-start recommendation.
      • 2023.Meta-Learning with Adaptive Weighted Loss for Imbalanced Cold-Start Recommendation.
    • 任务相关性性（Task relevance ）: Understanding task relevance is crucial for adapting warm global knowledge to cold-start scenarios, especially when cold users share similar preferences with warm users
      • 2021.Task-adaptive neural process for user cold-start recommendation.
      • 2021.Personalized adaptive meta learning for cold-start user preference prediction.
      • 2022.Task Similarity Aware Meta Learning for Cold-Start Recommendation
      • 2023.M2eu: Meta learning for cold-start recommendation via enhancing user preference estimation
  • 元嵌入（Meta-embedding initialization）：元嵌入初始化的目标在于生成预训练嵌入表示，从而加速冷启动用户及物品的拟合过程
    • 2017.Model-agnostic meta-learning for fast adaptation of deep networks.
    • 2019.Warm up cold-start advertisements: Improving ctr predictions via learning to learn id embeddings.
    • 2021.Learning to warm up cold item embeddings for cold-start recommendation with meta scaling and shifting networks.
  • 序列元学习（Sequential meta-learning）：序列元学习通过考虑用户交互的时间顺序，利用有限的历史行为序列来捕捉动态偏好
    • 2021 Sequential recommendation for cold-start users with meta transitional learning.
    • 2021.Cold-start sequential recommendation via meta learner.
    • 2022.A dynamic meta-learning model for time-sensitive cold-start recommendations.
    • 2022.Multimodal meta-learning for cold-start sequential recommendation.

GRAPH RELATIONS

• 交互图增强（Interaction Graph Enhancement）
  • 补充图关系（Supplementary Graph Relation）: 关键的问题在于为冷节点找到合适的生成边以及衡量边质量的方法
    • 2024.Content-based Graph Reconstruction for Cold-start Item Recommendation.
    • 2024.Mutual Information Assisted Graph Convolution Network for Cold-Start Recommendation.
    • 2023.Uncertainty-aware Consistency Learning for Cold-Start Item Recommendation.
  • 同质性网络关系（Homophily Network Relation ）：同质性假设的核心观点是认为网络中心节点与其相邻节点应具有相似的行为特征或标签信息
    • 2021.Cluster-based bandits: Fast cold-start for recommender system new users.
    • 2021.Learning graph meta embeddings for cold-start ads in click-through rate prediction
    • 2022.Socially-aware dual contrastive learning for cold-start recommendation.
    • 2023.A survey of graph neural network based recommendation in social networks
    • 2024.A survey of graph neural networks for social recommender systems.
    • 2024.Mitigating Extreme Cold Start in Graph-based RecSys through Re-ranking.
• 图关系扩展（Graph Relation Extension）
  • 异构图关系（Heterogeneous Graph Relation）：相较于交互图，此类方法通过扩展图中节点和边的类型已获取丰富的信息
    • 2020.A heterogeneous graph neural model for cold-start recommendation
    • 2021.Multi-view denoising graph auto-encoders on heterogeneous information networks for cold-start recommendation.
    • 2022.Gift: Graph-guided feature transfer for cold-start video click-through rate prediction.
    • 2021.Privileged graph distillation for cold start recommendation
    • 2023.User cold-start recommendation via inductive heterogeneous graph neural network
  • 属性图关系（Attributed Graph Relation）：
    • 2023.Multi-task item-attribute graph pre-training for strict cold-start item recommendation.
    • 2024.Warming Up Cold-Start CTR Prediction by Learning Item-Specific Feature Interactions.
  • 知识图关系（Knowledge Graph Relation）
    • 2021.Alleviating cold-start problems in recommendation through pseudo-labelling over knowledge graph.
    • 2022.Metakg: Meta-learning on knowledge graph for cold-start recommendation
    • 2024.Cold-Start Recommendation based on Knowledge Graph and Meta-Learning under Positive and Negative sampling.
• 图聚合器优化（Graph Aggregator Improvement）
  • 聚合域扩展（Expanding the aggregation scope）
    • 2023.Boosting Meta-Learning Cold-Start Recommendation with Graph Neural Network.
    • 2023.A Multi-strategy-based Pre-training Method for Cold-start Recommendation.
  • 聚合器增强（Augmenting the information aggregator）
    • 2021.Pre-training graph neural networks for cold-start users and items representation
    • 2024.Graph attention networks with adaptive neighbor graph aggregation for cold-start recommendation.

DOMAIN INFORMATION

• 域知识迁移（Domain Knowledge Transfer）
  • 嵌入映射（Embedding mapping）：
    • General Mapping
      • 2017.Cross-domain recommendation: An embedding and mapping approach
      • 2018.Transferable contextual bandit for cross-domain recommendation.
      • 2019.Semi-supervised learning for cross-domain recommendation to cold-start users.
      • 2020.Attentive-feature transfer based on mapping for cross-domain recommendation.
      • 2020.DCDIR: A deep cross-domain recommendation system for cold start users in insurance domain.
      • 2020.A heterogeneous information network based cross domain insurance recommendation system for cold start users.
    • Personalized Mapping
      • 2022.Fedcdr: federated cross-domain recommendation for privacy-preserving rating prediction
      • 2022.Personalized transfer of user preferences for cross-domain recommendation.
      • 2024.Vietoris-rips complex: A new direction for cross-domain cold-start recommendation.
  • 图关联（Graph connections）
    • 知识图（Knowledge Graph）：
      • 2023.REMIT: reinforced multi-interest transfer for cross-domain recommendation.
      • 2024.Domain-Oriented Knowledge Transfer for CrossDomain Recommendation
    • 混合关联（Hybrid Connections）
      • 2015.Social recommendation with cross-domain transferable knowledge
      • 2015.Improving top-n recommendation for cold-start users via cross-domain information
      • 2024.DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation.
  • 训练优化（Learning processes）
    • 训练技术（Training Techniques）
      • 2016.Learning informative priors from heterogeneous domains to improve recommendation in cold-start user domains.
      • 2018.A cross-domain recommendation mechanism for cold-start users based on partial least squares regression
    • Efficient Tuning
      • 2021.User-specific adaptive fine-tuning for cross-domain recommendations.
      • 2021.Transfer-meta framework for cross-domain recommendation to cold-start users.
      • 2023.Contrastive graph prompt-tuning for cross-domain recommendation.
      • 2024.CDRNP: Cross-Domain Recommendation to Cold-Start Users via Neural Process.
• 域分布对齐（Domain Distribution Alignment）
  • 协同过滤对齐（Collaborative Filtering Alignment）
    • Contrastive Alignment
      • 2022.Task-optimized user clustering based on mobile app usage for cold-start recommendations.
      • 2023.Cross-domain recommendation via user interest alignment.
      • 2024.User Distribution Mapping Modelling with Collaborative Filtering for Cross Domain Recommendation.
    • Latent-Dimension Alignment
      • 2022.Task-optimized user clustering based on mobile app usage for cold-start recommendations.
      • 2021.Low-dimensional alignment for cross-domain recommendation.
  • 辅助特征对齐（Auxiliary Feature Alignment）:
    • Stein-Path Alignment.
      • 2021.Leveraging distribution alignment via stein path for cross-domain cold-start recommendation.
      • 2023.Contrastive proxy kernel stein path alignment for cross-domain cold-start recommendation.
    • Contrastive Alignment
      • 2022.Contrastive learning for sequential recommendation.
• 域独立表征学习（Domain-Invariant Representation Learning）
  • 解耦表征（Disentangled Representation）
    • 对抗学习（Adversarial Learning）
      • 2019.RecSys-DAN: Discriminative adversarial networks for cross-domain recommender systems.
      • 2022.Cross-domain recommendation via adversarial adaptation.
      • 2024.Diff-MSR: A Diffusion Model Enhanced Paradigm for Cold-Start Multi-Scenario Recommendation.
    • 注意力机制（Attention Mechanism）
      • 2020.Internal and Contextual Attention Network for Cold-start Multi-channel Matching in Recommendation
      • 2024.Heterogeneous graph contrastive learning for cold start cross-domain recommendation
  • 融合表征（Fusing Representation）
    • 多视图学习（Multi-View Learning）
      • 2015.A multi-view deep learning approach for cross domain user modeling in recommendation systems.
      • 2024.A Dual Perspective Framework of Knowledgecorrelation for Cross-domain Recommendation.
    • 交换学习（Swapping Learning.）
      • 2020.Dual autoencoder network with swap reconstruction for cold-start recommendation.
      • 2020.CATN: Cross-domain recommendation for cold-start users via aspect transfer network.
      • 2022.Cross-domain recommendation to cold-start users via variational information bottleneck
    • 语义学习（Semantic Learning）：假设语义特征空间为领域不变的
      • 2018.A general cross-domain recommendation framework via Bayesian neural network
      • 2019.Preliminary investigation of alleviating user cold-start problem in e-commerce with deep cross-domain recommender system.
      • 2019.Deeply fusing reviews and contents for cold start users in cross-domain recommendation systems.
      • 2020.Zero-shot heterogeneous transfer learning from recommender systems to cold-start search retrieval

WORLD KNOWLEDGE FROM LARGE LANGUAGE MODELS

• LLM as the Recommender System

  • 提示策略（Prompting Strategy） : 该系列研究通过将推荐任务重构为自然语言处理问题，设计了多种提示策略来引导大语言模型（LLMs）生成准确的（系统）冷启动推荐

    • 直接提示（Direct Prompting）

      • 2022.Zero-shot recommendation as language modeling
      • 2023.Is chatgpt a good recommender? a preliminary study
      • 2023.Uncovering chatgpt’s capabilities in recommender systems.
      • 2023.Large language models as zero-shot conversational recommenders.
      • 2023.Large language models are competitive near cold-start recommenders for language-and item-based preferences.
      • 2024. Large language models are zero-shot rankers for recommender systems.
        • histories prompt
          • Sequential prompting:
            • “I’ve watched the following movies in the past in order: ’0. Multiplicity’, ’1.Jurassic Park’, . . .”.
          • Recency-focused prompting:
            • “I’ve watched the following movies in the past in order: ’0. Multiplicity’, ’1. Jurassic Park’, . . … Note that my most recently watched movie is Dead Presidents. . . .”.
          • In-context learning (ICL):
            • “ If I’ve watched the following movies in the past in order: ’0. Multiplicity’, ’1. Jurassic Park’, . . ., then you should recommend Dead Presidents to me and now that I’ve watched Dead Presidents, then . . .”.
        • candidate prompt
          • “Now there are 20 candidate movies that I can watch next: ’0. Sister Act’, ’1. Sunset Blvd’, . . .”.
        • Total Prompt
          • “ [pattern that contains sequential historical interactions H] [pattern that contains retrieved candidate items C] Please rank these movies by measuring the possibilities that I would like to watch next most, according to my watching history.”.
        • Findings :
          • 大语言模型（LLMs）往往难以有效感知给定的序列化交互历史中的顺序关系。通过采用特定设计的提示策略，可以激发LLMs对时序关系的感知能力，从而提升排序推荐性能
          • 大语言模型（LLMs）在进行排序推荐时容易受到位置偏差和流行度偏差的影响，但通过自助采样法或特殊设计的提示策略可有效缓解这些问题
          • 大语言模型（LLMs）在零样本推荐任务中表现优于现有方法，展现出强大的零样本排序能力
      • 2024. Large Language Models as Conversational Movie Recommenders: A User Study.
      • 2024. Could Small Language Models Serve as Recommenders? Towards Data-centric Cold-start Recommendation.

    • 多步提示（Muti-Step Prompting ）: 尽管直接提示法可用于零样本/少样本推荐任务，但其性能表现仍逊色于经过完整用户-物品交互数据训练的传统推荐方法

      • 2023. Zero-shot next-item recommendation using large pretrained language models.
        • (1) capture user preferences,
        • (2) select representative movies,
        • (3) recommend items based on these inputs.

      • 2024.Where to move next: Zero-shot generalization of llms for next poi recommendation

      • 2024.Large language models for next point-of-interest recommendation.

      • 2024. Taxonomy-Guided Zero-Shot Recommendations with LLMs.

      • 2024.Rec-GPT4V: Multimodal Recommendation with Large Vision-Language Models.

    • 检索增强（Retrieval-Augmented Recommendation）

      • 2023.Large language model augmented narrative driven recommendations.
      • 2023.Retrieval-augmented recommender system: Enhancing recommender systems with large language models.
      • 2024.RAGSys: Item-Cold-Start Recommender as RAG System.
      • 2024.Keyword-driven Retrieval-Augmented Large Language Models for Cold-start User Recommendations.
      • 2024. CoRAL: Collaborative Retrieval-Augmented Large Language Models Improve Long-tail Recommendation.
        • Retrieval Policy
        • Prompt

  • 模型调优（Model Tuning）

    • 指令调优（Instruction Tuning）：本质上是将推荐任务转化为文本到文本的生成过程。该框架通过下一词元预测训练，使模型学会用自然语言解析用户历史行为和物品元数据
      • 指令调优一般分为四步 :
        • Define a task and articulate a “Task Instruction” using natural language
        • Formulate and construct the input and output of the task in natural language, denoted as “Task Input” and “Task Output”.
        • Integrate the “Task Instruction” and “Task Input” together to form the “Instruction Input”, and take the “Task Output” as the corresponding “Instruction Output”
        • Instruction tuning on LLMs based on the formatted pairs of “Instruction Input” and “Instruction Output”.
    • Papers :
      • 2022.Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5).
      • 2023.Leveraging large language models for pre-trained recommender systems.
      • 2023.Tallrec: An effective and efficient tuning framework to align large language model with recommendation.
      • 2023.Collm: Integrating collaborative embeddings into large language models for recommendation.
      • 2024.Large language models meet collaborative filtering: an efficient all-round LLM-based recommender system.
      • 2024.Text-like Encoding of Collaborative Information in Large Language Models for Recommendation.
      • 2024.Bridging items and language: A transition paradigm for large language model-based recommendation.
      • 2024. Idgenrec: Llm-recsys alignment with textual id learning.
        • 步骤:
          • LLM-based ID Generator
          • Prompt Template :
            • User [user_ID] has purchased items [item_ID], [item_ID], …, [item_ID]; predict the next possible item to be bought by the user
          • Instruction Tuning LLM-based Base Recommender
    • 模型调优（Fine-Tuning）：
      • 2023.Prompt distillation for efficient llm-based recommendation
      • 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation.
      • 2024.NoteLLM-2: Multimodal Large Representation Models for Recommendation.
      • 2024.Exploring User Retrieval Integration towards Large Language Models for Cross-Domain Sequential Recommendation
      • 2024.XRec: Large Language Models for Explainable Recommendation.

• LLM as the Knowledge Enhancer

  • 基于LLM的表征增强（LLM for Representation Enhancement）
    • 多模态增强（Modality-Enhanced Representation）
      • 2024.General Item Representation Learning for Cold-start Content Recommendations.
      • 2024.EasyRec: Simple yet Effective Language Models for Recommendation.
      • 2024. Enhancing sequential recommendation via llm-based semantic embedding learning.
        • 设定：
          • sequential recommendation modeling
          • each item is associated with a unique ID and several textual attributes
      • 方案
        • two-stage training process
        • semantically aligned embedding learning
          • the goal of Projector is to elicit the text sequence 𝜏𝑖 from the LLM, given the projected embedding 𝑥𝑖 as input to it
            • LLM(<23>) = Brand
            • LLM(<23, Brand>) => BrandA
            • LLM(<23, Brand, BrandA>) = Category
            • …
        • model-agnostic sequential recommender training.
          • （传统序列建模，略…）
    • 域增强（Domain-Enhanced Representation）
      • 2023. One model for all: Large language models are domain-agnostic recommendation systems.
        • Multi-domain Sequential Recommendation
      • 2023.An Unified Search and Recommendation Foundation Model for Cold-Start Scenario.
  • 基于LLM的关系增强（LLM for Relation Augmentation）
    • 行为模拟（Behavior Simulation）
      • 2024.Large Language Model Interaction Simulator for Cold-Start Item Recommendation.
      • 2024. Large Language Models as Data Augmenters for Cold-Start Item Recommendation.
        • steps:
          • Augmented Data Generation
          • Pairwise Comparison Loss
            • \mathcal{L}{aug} = -\sum{(u, pos, neg)}ln\sigma(\hat{y}{u,pos} - \hat{y}{u.neg})
    • 外部关联补充（External Relation Supplement）：利用大模型学习到的世界知识与常识信息来建立用户-物品交互关联
      • 2024.Comprehending Knowledge Graphs with Large Language Models for Recommender Systems.
      • 2024. Common Sense Enhanced Knowledge-based Recommendation with Large Language Model.
        • Steps:
          • Common Sense-based KG(Knowledge Graph) construction ：大模型的利用点
            • Entity :
              • Categories
            • Relation :
              • complementarity : refers to the scenario where two items can be used together, such as a phone and a phone case.
              • substitutes : imply that two items can replace each other and serve similar functions, for example, a tablet and an e-reader.
            • 关系生成
              • Prompt :
                • “I have bought an item, whose category is {item category}. Please recommend 10 categories of complementary/substitutable items. Output format: One category one line, without any explanations.”
          • Recommendation with Common Sense-based KG