└── README.md
/README.md:
--------------------------------------------------------------------------------
1 | # LLM4REC
2 | ## OVERVIEW
3 | 1. LLMs enhance Recommendation
4 | - Feature Engineering
5 | - data augmentation
6 | - generate open-world knowledge for user/item
7 | - generate interaction data
8 | - data condense
9 | - feature selection
10 | - feature imputation
11 | - Feature Encoder
12 | - encode text information
13 | - encode id information
14 |
15 | 2. LLMs as Recommenders
16 | - prompt learning
17 | - instruction tuning
18 | - reinforce learning
19 | - knowledge distillation
20 | - Pipeline Controller
21 | - pipeline design
22 | - CoT, ToT, SI
23 | - Incremental Learning
24 |
25 | 3. Other Related work
26 | - Self-distillation in LLM
27 | - DPO in LLM
28 | - LLM4CTR
29 |
30 | ## 1. LLMs enhance Recommendation
31 |
32 | ### Feature Engineering
33 | | Title | Model | Time | Motivation | Discription |
34 | |:-------:|:-------:|:-------:|:-------:|:-------:|
35 | | Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models | CIKGRec | AAAI25 | 结构化LLM中用户侧的世界知识，增强知识感知的基于图的推荐算法 |  |
36 | | Towards Open-World Recommendation with Knowledge Augmentation from Large Language Models | KAR | RecSys24 | 利用LLM的open-world knowledge扩充用户和物品的信息 ||
37 | | A First Look at LLM-Powered Generative News Recommendation | ONCE(GENRE+DIRE) | arXiv23 | 对于开源LLM，利用它们作为特征编码器。对于闭源LLM，使用提示丰富训练数据 | |
38 | | LLMRec: Large Language Models with Graph Augmentation for Recommendation | LLMRec | WSDM24 | 利用LLM进行图数据增强，从item candidates中选出liked item和disliked item | |
39 | | Integrating Large Language Models into Recommendation via Mutual Augmentation and Adaptive Aggregation | Llama4Rec | arXiv24 | 由mutual augmentation和adaptive aggregation组成。mutual augmentation包括data增强和prompt增强。 | |
40 | | Data-efficient Fine-tuning for LLM-based Recommendation | DEALRec | SIGIR24 | 设计influence score和effort score，对LLM4REC进行数据蒸馏，挑选出有influential的samples ||
41 | | Distillation is All You Need for Practically Using Different Pre-trained Recommendation Models |PRM-KD |arXiv24|利用了不同类型的预训练推荐模型作为教师模型，提取in-batch negative item scores进行联合知识蒸馏||
42 | | CoRAL: Collaborative Retrieval-Augmented Large Language Models Improve Long-tail Recommendation | CoRAL | KDD24 | 通过强化学习，将协同信息以prompt的形式增强LLM，实现对于Long-tail Recommendation推荐性能的改进||
43 | | Harnessing Large Language Models for Text-Rich Sequential Recommendation | |WWW24|关注LLM4REC的数据压缩问题，先将用户历史交互分片，然后用LLM总结每个分片的内容，最后设计prompt将总结后的user偏好、最近user交互和candidate items结合在一起||
44 | | Large Language Models Enhanced Collaborative Filtering |LLM-CF|CIKM24|通过ICL和COT，将LLM的world knowledge和reasoning capabilities蒸馏到collaborative filtering||
45 | | Optimization Methods for Personalizing Large Language Models through Retrieval Augmentation | | SIGIR24 | LLMs不能根据其用户的背景和历史偏好定制其生成的输出,通过强化学习+知识蒸馏选择最能增强LLM的个人信息 |  |
46 | | Large Language Models for Next Point-of-Interest Recommendation | | SIGIR24 | 现有的next POI方法侧重于短轨迹和冷启动问题（数据量少且轨迹短的用户），没有充分探索丰富的LBSN的数据,可以使用LLM的自然语言理解能力，来处理所有类型的LBSN数据并更好地使用上下文信息 |  |
47 |
48 |
49 | ### Feature Encoder
50 | | Title | Model | Time | Motivation | Discription |
51 | |:-------:|:-------:|:-------:|:-------:|:-------:|
52 | | U-BERT: Pre-training user representations for improved recommendation | U-BERT | AAAI21 | 早期的工作，主要使用BERT编码评论文本 | |
53 | | Towards universal sequence representation learning for recommender systems | UniSRec | KDD22 | 用BERT对item text信息进行编码，使用了parametric whitening | |
54 | | Learning vector-quantized item representation for transferable sequential recommenders | VQ-Rec | WWW23 | 首先将文本映射到一个离散索引向量（称为item code ）中，然后使用这些索引来查找code embedding table进行编码 | |
55 | | Recommender Systems with Generative Retrieval | TIGER | NIPS23 |使用LLM编码有意义的item ID，直接预测candidate IDs，进行端到端的generative retrieval | |
56 | | Representation Learning with Large Language Models for Recommendation | RLMRec | WWW24 | 通过两次对比学习，对齐LLM编码的语义特征和传统方法的协同特征 | |
57 | | Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recommendation | ReLLa | WWW24 | CTR问题，LLM对于长的序列效果不佳；本文根据target item从长序列中选择相似的部分item作为序列；item的embedding通过LLM对text信息进行构建 | 
|
58 | | Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors | BAHE | SIGIR24 short paper | 长序列LLM推理开销大。本文思路是固定LLM的浅层参数，预先存储一些原子交互的LLM的浅层特征，后续直接查表 | 
|
59 | | Large Language Models Augmented Rating Prediction in Recommender System | LLM-TRSR | ICASSP24 | ensemble LLM_Rec和传统Rec的输出 | 
|
60 | | Enhancing Content-based Recommendation via Large Language Model | LOID | CIKM24 short paper | 不同domain的content语义信息之间可能有gap；同时利用LLM和传统RS的信息，提出一种ID和content信息align的范式。用ID embedding作为key提取text embedding序列当中的信息 | 
61 | | Aligning Large Language Models with Recommendation Knowledge | | arXiv24 | 将推荐领域的一些知识，例如MIM和BPR，通过prompt的形式将其传输给LLM | 
|
62 | | The Elephant in the Room: Rethinking the Usage of Pre-trained Language Model in Sequential Recommendation | Elephant in the Room | RecSys24 | 序列推荐的大模型的attention层的大部分参数都没有被使用，参数存在大量的冗余。本文将LLM学到的item embedding作为SASRec的初始化，然后再训练SASRec | 
|
63 | | Demystifying Embedding Spaces using Large Language Models | | ICLR24 | 用LLM对item的embedding空间进行解释，包括未在训练数据中出现过的item | 
|
64 |
65 |
66 | ## 2. LLMs serve as Recommenders
67 | ### Scoring/Ranking
68 | | Title | Model | Time | Motivation | Discription |
69 | |:-------:|:-------:|:-------:|:-------:|:-------:|
70 | | Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5) | P5 | RecSys22 | 针对不同任务设计了多个prompts，并且使用推荐数据集重新进行预训练，最终用于解决zero-shot的推荐问题 | |
71 | | Text Is All You Need: Learning Language Representations for Sequential Recommendation | RecFormer | KDD23 | 将键值对展开为类似句子的prompt ，利用LongFormer训练，输出用户交互序列（兴趣）的表征。然后结合对比学习，进行最后的推荐 | |
72 | | Recommendation as instruction following: A large language model empowered recommendation approach | InstructRec | arXiv23 | 采用instruction tuning，将主动的用户指令和被动的交互信息按照一定格式组织成指令，引导LLM完成多任务推荐场景 | |
73 | | A bi-step grounding paradigm for large language models in recommendation systems | BIGRec | arXiv23 | 针对grounding问题，采用instruction-tuning，实现“Grounding Language Space to Recommendation Space” | |
74 | | A Multi-facet Paradigm to Bridge Large Language Model and Recommendation | TransRec | arXiv23 | 在Item indexing上，将ID, title和attribute都当成Item的facet；在generation grounding上，：将生成的identifiers与in-corpus 的每个item的identifiers取交集选出items | |
75 | | CoLLM: Integrating Collaborative Embeddings into Large Language Models for Recommendation | CoLLM | arXiv23 | 将传统模型捕获的协作放到LLM的prompt中，并将其映射到最终的embedding空间 | |
76 | | LlamaRec: Two-Stage Recommendation using Large Language Models for Ranking | LlamaRec | CIKM23 | 一般LLM生成推荐结果的推理成本很高，并且要进一步Grounding。LlamaRec利用一个verbalizer ，将LLM head的输出(即所有tokens的分数)转换为候选items的排名分数 | |
77 | | Large language models are zero-shot rankers for recommender systems | | arXiv23 | 利用LLM对候选物品集合进行zero-shot排序 | |
78 | | Language models as recommender systems: Evaluations and limitations | LMRecSys | NeurIPS21 | 采用Prompt tuning的方法，将要预测的物品拆分成多个token，由LLM输出每个token的分布，最终进行推荐 | |
79 | | Prompt learning for news recommendation | Prompt4NR | SIGIR23 | 设计离散、连续、混合提示模板，以及它们对应的答案空间。使用prompt ensembling组合效果最好的一组prompt模板 | |
80 | | Prompt distillation for efficient llm-based recommendation | POD | CIKM23 | 通过prompt learning学习作为前缀的连续prompt，将离散prompt信息蒸馏到连续prompt | |
81 | | Large Language Models as Zero-Shot Conversational Recommenders | | CIKM23 | 使用具有代表性的大型语言模型在Zero-Shot下对会话推荐任务进行实证研究 | |
82 | | Leveraging Large Language Models (LLMs) to Empower Training-Free Dataset Condensation for Content-Based Recommendation | | arXiv23 | 对推荐数据进行蒸馏，设计prompt，用LLM压缩item的信息、提取user偏好，聚类并计算距离选择top-m的user，并产生交互数据 | |
83 | | Is ChatGPT Good at Search? Investigating Large Language Models as Re-Ranking Agents | | arXiv23 | 利用GPT模型进行文本排序任务，将GPT模型的标注结果用于模型蒸馏 | |
84 | | LLaRA: Aligning Large Language Models with Sequential Recommenders | LLaRA | arXiv23 | 在prompt中采用文本表征+传统模型学习的混合表征 ||
85 | | Collaborative Contextualization: Bridging the Gap between Collaborative Filtering and Pre-trained Language Model | CollabContext | arXiv23 | 利用LLM学习到的文本表征和传统模型表征进行双向蒸馏 ||
86 | |Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation|LC-Rec | ICDE24 |对于item索引，设计了一种语义映射方法，可以为item分配有意义且不冲突的id，同时提出了一系列特别设计的tuning任务，迫使llm深度整合语言和协同过滤语义||
87 | | Collaborative Large Language Model for Recommender Systems| CLLM4Rec| WWW24 | 为了减少自然语言和推荐语义的gap，本文为user和item扩充词表使其与唯一的token绑定，并引入协同信号进行训练扩充的token的embedding | |
88 | | Play to Your Strengths: Collaborative Intelligence of Conventional Recommender Models and Large Language Models | Play to Your Strength | arxiv24 | CTR task；由于LLM inference时间过长，且传统RS和LLM RS擅长不同的数据，本文考虑对不同数据分别使用传统RS和LLM进行推荐。方法是将传统RS confidence低的sample丢给LLM RS判断 | 
|
89 | |GPT4Rec: A generative framework for personalized recommendation and user interests interpretation|GPT4Rec| arxiv23 | 用GPT2根据历史交互产生query，在BM25中检索item |  |
90 | | Unsupervised large Language Model Alignment for Information Retrieval via Contrastive Feedback | | SIGIR24 | LLMs产生的responses不能捕捉内容相似的document之间的区别,设计group-wise的方法产生反馈信号，用无监督学习+强化学习，使LLMs产生context-specific的responses|  |
91 | | RDRec: Rationale Distillation for LLM-based Recommendation | RDRec | arXiv24 | 现在的LLM4REC很少关注user产生interaction背后的rationale；让LLM通过prompt从review中提取user preference和item attribute，然后利用小LM进行蒸馏 |  |
92 |
93 | ### Pipline Controller
94 | | Title | Model | Time | Motivation | Discription |
95 | |:-------:|:-------:|:-------:|:-------:|:-------:|
96 | | Recmind: Large language model powered agent for recommendation | Recmind | arXiv23 | 由LLM驱动的推荐Agent，可以推理、互动、记忆，提供精确的个性化推荐 | |
97 | | Can Small Language Models be Good Reasoners for Sequential Recommendation? | SLIM | WWW24 | 将大的LLM的逐步推理能力蒸馏到小的LLM中 |  |
98 | |Preliminary Study on Incremental Learning for Large Language Model-based Recommender Systems| |arXiv23|实验发现full retraining and fine-tuning增量学习都没有显著提高LLM4Rec的性能，设计long term lora(freeze)和short term lora(hot)分别关注user长短期偏好||
99 | |Scaling Law of Large Sequential Recommendation Models| | arXiv23 |实验发现，扩大模型大小可以极大地提高具有挑战性的推荐任务上(如冷启动、鲁棒性、长期偏好)的性能||
100 |
101 |
102 | ## 3. Other Related work
103 | ### Self-distillation in LLM
104 | | Title | Model | Time | Motivation | Discription |
105 | |:-------:|:-------:|:-------:|:-------:|:-------:|
106 | | SELF-INSTRUCT: Aligning Language Models with Self-Generated Instructions | SELF-INSTRUCT | arXiv22 | 对于指令微调，人类编写的指令数据开销大，多样性有限、不能推广到广泛的场景；可以让LLM自己产生指令| |
107 | | Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision | SELF-ALIGN | NeurIPS23 | 人类注释的监督微调(SFT)和来自人类反馈的强化学习(RLHF)有成本高，可靠性、多样性参差不齐等问题;可以将原则驱动推理和LLM的生成能力结合起来，在最少的人类监督下实现人工智能agent的自对齐| |
108 | | RLCD: REINFORCEMENT LEARNING FROM CONTRASTIVE DISTILLATION FOR LM ALIGNMENT| RLCD | ICLR24 | 设计对比学习，在不使用人类反馈的情况下，使语言模型遵循自然语言表达的原则的方法产生指令(从模型输出中创建偏好对，一个旨在鼓励遵循给定原则的积极提示，另一个旨在鼓励违反原则的消极提示) | |
109 | | Impossible Distillation for Paraphrasing and Summarization: How to Make High-quality Lemonade out of Small, Low-quality Models| | arXiv23 | 从低质量教师模型（本身不能执行某些特定任务任务的模型）提取出高质量的数据集和模型,最后，学生LM通过自我蒸馏进一步完善（在自己的高质量数据上进行训练）| |
110 | | LARGE LANGUAGE MODELS CAN SELF-IMPROVE | | arXiv23 | 微调LLM需要大量有监督数据，而人类的反思不需要外部输入；可以让LLM通过unlabeled数据进行反思;通过Chain-of-Thought prompting 和 self-consistency 让LLM产生“high-confidence” 的回答 |  |
111 | | Reinforced Self-Training (ReST) for Language Modeling | ReST | arXiv23 | RLHF通过将LLM和人类偏好对齐来提升LLM的能力，它采用的在线训练策略在处理新的样本时开销大；可以采用离线强化学习来解决这个问题（时间问题）；离线的强化学习的质量很大程度上取决于数据集的质量，需要得到高质量的离线数据集（提升有效性）|  |
112 | | Self-Rewarding Language Models | | arXiv24 | 目前的RLHF根据人类偏好来训练奖励模型，这受到人类表现水平的显示；其次这些冻结的奖励模型无法在LLM训练的过程中学习改进；需要让LLM自动修改奖励函数，并且在训练的过程中自动改进 | |
113 | | Baize: An Open-Source Chat Model with Parameter-Efficient Tuning on Self-Chat Data | Baize | arXiv23 | 目前具备强大能力的聊天模型如ChatGPT访问经常受限（只能通过API访问），希望能够训练一个能力接近ChatGPT的开源模型；为了让开源LLM的聊天能力接近ChatGPT，需要为开源LLM提供高质量的训练数据；通过利用ChatGPT与自己进行对话，可以自动生成高质量的多回合聊天语料库；提出带有反馈的自蒸馏，以进一步提高带有ChatGPT反馈的Baize模型的性能| |
114 | | STaR: Self-Taught Reasoner Bootstrapping Reasoning With Reasoning | STaR | arXiv22 | 思维链能够提升LLM在复杂推理场景的表现，但是这类方法有个缺点：它们要么需要大量的思维链数据，开销很大；要么只使用少量的思维连数据，损失了一部分推理的能力；希望LLM学习自己生成的rationale来提升推理能力，但自己生成的rationale可能是错误的answer，需要修正|  |
115 | | Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning | | arXiv24 | 为特定任务对LLM进行微调通常会面临一个挑战：平衡对特定任务性能和对一般任务指令的遵循能力；LLM重写特定任务的response，来减少两种分布之间的gap |  |
116 |
117 |
118 | ### Direct Preference Optimization in LLM (DPO)
119 | | Title | Model | Time | Motivation | Discription/Loss Fuction |
120 | |:-------:|:-------:|:-------:|:-------:|:-------:|
121 | | Direct Preference Optimization: Your Language Model is Secretly a Reward Model | DPO | NeurIPS24 | 省去RLHF对于reward model的构建，直接针对偏好数据进行模型的优化 |  |
122 | | Statistical Rejection Sampling Improves Preference Optimization | RSO | ICLR24 | 提出DPO的偏好数据并非采样自最优策略，引入显式的reward模型和统计拒绝采样使产生自SFT模型的数据分布可以拟合最优模型的数据分布 |  |
123 | | KTO: Model Alignment as Prospect Theoretic Optimization | KTO | arXiv24 | 将DPO修正为针对label数据而非偏好数据对的优化 ||
124 | | Curry-DPO: Enhancing Alignment using Curriculum Learning & Ranked Preferences | Curry-DPO | arXiv24 | 对于同个prompt的多个response，按照reward的差值构造pairwise数据对，再利用课程学习由易到难进行训练 | |
125 | | LiPO: Listwise Preference Optimization through Learning-to-Rank | LiPO | arXiv24 | 修正DPO的loss，直接对listwise数据进行优化 | |
126 | | ULMA: Unified Language Model Alignment with Human Demonstration and Point-wise Preference | ULMA| arXiv23 | 修正DPO的loss，直接对pointwise数据进行优化 ||
127 | | Reinforcement Learning from Human Feedback with Active Queries | ADPO | arXiv24 | active-learning的范式，去除reward差值小的数据对 | |
128 | | RS-DPO: A Hybrid Rejection Sampling and Direct Preference Optimization Method for Alignment of Large Language Models| RS-DPO | arXiv24 | 引入显式的reward模型，使用拒绝统计采样，去除reward差值小的数据对，提高样本利用效率 | |
129 | | Direct Preference Optimization with an Offset| ODPO| arXiv24 | 引入偏移值来表示偏好数据集中喜欢相对于不喜欢的程度 | |
130 | | BRAIN: Bayesian Reward-conditioned Amortized INference for natural language generation from feedback | BRAIN | arXiv24 | 重新引入reward模型表示偏好数据集中喜欢相对于不喜欢的程度 | |
131 | | D2PO: Discriminator-Guided DPO with Response Evaluation Models | D2PO | arXiv24 | online训练方式，同时训练一个reward模型，在训练过程中迭代地由当前模型和reward模型产生新样本| |
132 | |Learn Your Reference Model for Real Good Alignment|TR-DPO|arXiv24|使用soft和hard两种更新方式，在训练期间更新reference model||
133 | |sDPO: Don’t Use Your Data All at Once|sDPO|arXiv24|分批利用训练数据集，并在训练过程中更ref模型||
134 | |Direct Language Model Alignment from Online AI Feedback|OAIF|arXiv24|利用更优模型在训练过程中产生新的偏好数据对||
135 | | A General Theoretical Paradigm to Understand Learning from Human Preferences | IPO | PMLR24 | 在DPO loss上加了一个正则化项，避免训练时快速overfitting| |  |
136 | | Provably Robust DPO: Aligning Language Models with Noisy Feedback | rDPO | arXiv24 | 修正DPO的loss,使其对偏好数据概率翻转鲁棒 |  |
137 | | Zephyr: Direct Distillation of LM Alignment | Zephyr | arXiv23 | 利用大模型（GPT4）生成偏好数据，再使用DPO对7B模型进行微调 |  |
138 |
139 | ### LLM4CTR
140 | | Title | Model | Time | Description |
141 | |:-------:|:-------:|:-------:|:-------:|
142 | | CTR-BERT: Cost-effective knowledge distillation for billion-parameter teacher models | CTR-BERT | NIPS WS'21 | CTR-BERT 提出了一种成本效益的知识蒸馏方法，用于十亿参数教师模型。 |
143 | | DCAF-BERT: A Distilled Cachable Adaptable Factorized Model For Improved Ads CTR Prediction | DCAF-BERT | WWW'22 | DCAF-BERT 提出了一种经过蒸馏的可缓存可适应因式化模型，用于提高广告点击率预测的准确性。 |
144 | | Learning Supplementary NLP Features for CTR Prediction in Sponsored Search | - | KDD'22 | 为了在赞助搜索中进行点击率预测，该研究探索了学习补充自然语言处理特征的方法。 |
145 | | Practice on Effectively Extracting NLP Features for Click-Through Rate Prediction | - | CIKM'23 | 通过实践，研究了有效提取自然语言处理特征用于点击率预测的方法。 |
146 | | BERT4CTR: An Efficient Framework to Combine Pre-trained Language Model with Non-textual Features for CTR Prediction | BERT4CTR | KDD'23 | BERT4CTR 提出了一种高效的框架，将预训练语言模型与非文本特征结合，用于点击率预测。 |
147 | | M6-rec: Generative pretrained language models are open-ended recommender systems | M6-rec | arxiv'22 | M6-rec 提出了一种生成式预训练语言模型，用作开放式推荐系统。 |
148 | | Ctrl: Connect tabular and language model for ctr prediction | Ctrl | arxiv'23 | Ctrl 提出了一种连接表格数据和语言模型用于点击率预测的方法。 |
149 | | FLIP: Towards Fine-grained Alignment between ID-based Models and Pretrained Language Models for CTR Prediction | FLIP | arxiv'23 | FLIP 旨在实现基于ID的模型和预训练语言模型之间的细粒度对齐，用于点击率预测。 |
150 | | TBIN: Modeling Long Textual Behavior Data for CTR Prediction | TBIN | arxiv'23 | TBIN 提出了一种用于点击率预测的长文本行为数据建模方法。 |
151 | | An Unified Search and Recommendation Foundation Model for Cold-Start Scenario | - | CIKM'23 | 为冷启动场景提出了一个统一的搜索和推荐基础模型。 |
152 | | A Unified Framework for Multi-Domain CTR Prediction via Large Language Models | - | arxiv'23 | 提出了一个通过大型语言模型进行多领域点击率预测的统一框架。 |
153 | | UFIN: Universal Feature Interaction Network for Multi-Domain Click-Through Rate Prediction | UFIN | arxiv'23 | UFIN 提出了一种通用特征交互网络，用于多领域的点击率预测。 |
154 | | ClickPrompt: CTR Models are Strong Prompt Generators for Adapting Language Models to CTR Prediction | ClickPrompt | WWW'24 | ClickPrompt 提出了点击率模型作为强大提示生成器，用于调整语言模型以进行点击率预测。 |
155 | | PRINT: Personalized Relevance Incentive Network for CTR Prediction in Sponsored Search | PRINT | WWW'24 | PRINT 提出了一种个性化相关性激励网络，用于赞助搜索中的点击率预测。 |
156 | | Breaking the Length Barrier: LLM-Enhanced CTR Prediction in Long Textual User Behaviors | - | arxiv'24 | 提出一种在长文本用户行为中增强点击率预测的LLM方法。 |
157 | | KELLMRec: Knowledge-Enhanced Large Language Models for Recommendation | KELLMRec | arxiv'24 | KELLMRec 提出了一种增强知识的大型语言模型用于推荐任务。 |
158 | | Enhancing sequential recommendation via llm-based semantic embedding learning | - | WWW'24 | 通过基于LLM的语义嵌入学习来增强顺序推荐任务。 |
159 | | Heterogeneous knowledge fusion: A novel approach for personalized recommendation via llm | - | Recsys'23 | 通过异构知识融合，提出了一种通过LLM进行个性化推荐的新方法。 |
160 | | Play to Your Strengths: Collaborative Intelligence of Conventional Recommender Models and Large Language Models | - | arxiv'24 | 利用传统推荐模型和大型语言模型的协同智能来发挥各自优势。 |
161 | | Generative Explore-Exploit: Training-free Optimization of Generative Recommender Systems using LLM Optimizers | - | arxiv'24 | 通过LLM优化器实现生成式推荐系统的无训练优化，实现探索-利用策略。 |
162 | #### Feature Selection
163 | | Title | Model | Time | Description |
164 | |:-------:|:-------:|:-------:|:-------:|
165 | | ICE-SEARCH: A Language Model-Driven Feature Selection Approach | ICE-SEARCH | arXiv'24 | ICE-SEARCH 提出了一种基于语言模型的特征选择方法。 |
166 | | Large Language Model Pruning | - | arXiv'24 | Model Pruning | 针对大型语言模型的剪枝技术研究。 |
167 | | Dynamic and Adaptive Feature Generation with LLM | - | arXiv'24 | 利用LLM进行动态和自适应特征生成。 |
168 |
--------------------------------------------------------------------------------