1. 通过多层次交互式对比学习改进知识感知推荐

Title: Improving Knowledge-aware Recommendation with Multi-level Interactive Contrastive Learning

Published: 2022-08-22

Url: http://arxiv.org/abs/2208.10061v1

Authors: Ding Zou,Wei Wei,Ziyang Wang,Xian-Ling Mao,Feida Zhu,Rui Fang,Dangyang Chen

将知识图谱（KG）整合到推荐系统中已经引起了相当大的关注。最近，知识感知推荐（KGR）的技术趋势是开发基于图神经网络（GNN）的端到端模型。然而，极其稀疏的用户-项目交互显着降低了基于 GNN 的模型的性能，因为：1）稀疏交互意味着监督信号不足并限制了基于 GNN 的监督模型；2）稀疏交互（CFpart）和冗余KG事实（KG part）的组合导致信息利用不平衡。此外，GNN 范式聚合了局部邻居进行节点表示学习，同时忽略了非局部 KG 事实，使得知识提取不足。受最近对比学习在从数据本身挖掘监督信号方面取得成功的启发，本文重点探索 KGR 中的对比学习，并提出了一种新颖的多层次交互式对比学习机制。与传统的对比学习方法对两个生成的图视图的节点进行对比，交互式对比机制通过对比图中不同部分的层来进行分层自监督学习，这也是一种“交互”动作。具体来说，我们首先为 KG 中的用户/项目构建本地和非本地图，为 KGR 探索更多的 KG 事实。然后在每个图中执行图内级别的交互式对比学习，对比 CF 和 KG 部分的层，以获得更一致的信息利用。此外，在局部和非局部图之间进行图间级交互式对比学习，以充分、连贯地提取非局部KG信号。在三个基准数据集上进行的大量实验表明，我们提出的方法优于最先进的方法。

Incorporating Knowledge Graphs (KG) into recommeder system has attractedconsiderable attention. Recently, the technical trend of Knowledge-awareRecommendation (KGR) is to develop end-to-end models based on graph neuralnetworks (GNNs). However, the extremely sparse user-item interactionssignificantly degrade the performance of the GNN-based models, as: 1) thesparse interaction, means inadequate supervision signals and limits thesupervised GNN-based models; 2) the combination of sparse interactions (CFpart) and redundant KG facts (KG part) results in an unbalanced informationutilization. Besides, the GNN paradigm aggregates local neighbors for noderepresentation learning, while ignoring the non-local KG facts and making theknowledge extraction insufficient. Inspired by the recent success ofcontrastive learning in mining supervised signals from data itself, in thispaper, we focus on exploring contrastive learning in KGR and propose a novelmulti-level interactive contrastive learning mechanism. Different fromtraditional contrastive learning methods which contrast nodes of two generatedgraph views, interactive contrastive mechanism conducts layer-wiseself-supervised learning by contrasting layers of different parts withingraphs, which is also an "interaction" action. Specifically, we first constructlocal and non-local graphs for user/item in KG, exploring more KG facts forKGR. Then an intra-graph level interactive contrastive learning is performedwithin each graph, which contrasts layers of the CF and KG parts, for moreconsistent information leveraging. Besides, an inter-graph level interactivecontrastive learning is performed between the local and non-local graphs, forsufficiently and coherently extracting non-local KG signals. Extensiveexperiments conducted on three benchmark datasets show the superior performanceof our proposed method over the state-of-the-arts.

2. 面向社交推荐的解耦式对比学习

Title: Disentangled Contrastive Learning for Social Recommendation

Published: 2022-08-18

Url: http://arxiv.org/abs/2208.08723v1

Authors: Jiahao Wu,Wenqi Fan,Jingfan Chen,Shengcai Liu,Qing Li,Ke Tang

社交推荐利用社交关系来增强推荐的表征学习。大多数社交推荐模型统一了用户-项目交互（协作领域）和社交关系（社交领域）的用户表示。然而，这种方法可能无法对两个域中的用户异构行为模式进行建模，从而损害用户表示的表达能力。在这项工作中，为了解决这种限制，我们提出了一种新颖的 Disentangled 对比学习框架，用于社会推荐 DcRec。更具体地说，我们建议从项目和社交领域学习解开的用户表示。此外，分离对比学习旨在执行分离的用户表示之间的知识转移，以进行社会推荐。在各种现实世界数据集上的综合实验证明了我们提出的模型的优越性。

Social recommendations utilize social relations to enhance the representationlearning for recommendations. Most social recommendation models unify userrepresentations for the user-item interactions (collaborative domain) andsocial relations (social domain). However, such an approach may fail to modelthe users heterogeneous behavior patterns in two domains, impairing theexpressiveness of user representations. In this work, to address suchlimitation, we propose a novel Disentangled contrastive learning framework forsocial Recommendations DcRec. More specifically, we propose to learndisentangled users representations from the item and social domains. Moreover,disentangled contrastive learning is designed to perform knowledge transferbetween disentangled users representations for social recommendations.Comprehensive experiments on various real-world datasets demonstrate thesuperiority of our proposed model.

3. 用于序列推荐的双向 Transformer 对比学习

Title: Contrastive Learning with Bidirectional Transformers for Sequential Recommendation

Published: 2022-08-17

Url: http://arxiv.org/abs/2208.03895v3

Authors: Hanwen Du,Hui Shi,Pengpeng Zhao,Deqing Wang,Victor S. Sheng,Yanchi Liu,Guanfeng Liu,Lei Zhao

使用基于 Transformer 的序列编码器的对比学习在序列推荐方面取得了优势。它最大化了共享相似语义的成对序列增强之间的协议。然而，现有的顺序推荐中的对比学习方法主要集中在从左到右的单向 Transformer 作为基础编码器，由于用户行为可能不是刚性的从左到右序列，因此对于顺序推荐来说不是最优的。为了解决这个问题，我们提出了一个名为 \textbf{C} 对比学习的新框架，它使用 \textbf{Bi}directional\textbf{T} 转换器进行顺序推荐 (\textbf{CBiT})。具体来说，我们首先将滑动窗口技术应用于双向 Transformers 中的长用户序列，这允许对用户序列进行更细粒度的划分。然后我们结合完形填空任务掩码和dropout掩码生成高质量的正样本并进行多对对比学习，与普通的一对对比学习相比，表现出更好的性能和适应性。此外，我们引入了一种新颖的动态损失重加权策略来平衡完形填空任务损失和对比损失。在三个公共基准数据集上的实验结果表明，我们的模型在顺序推荐方面优于最先进的模型。

Contrastive learning with Transformer-based sequence encoder has gainedpredominance for sequential recommendation. It maximizes the agreements betweenpaired sequence augmentations that share similar semantics. However, existingcontrastive learning approaches in sequential recommendation mainly center uponleft-to-right unidirectional Transformers as base encoders, which aresuboptimal for sequential recommendation because user behaviors may not be arigid left-to-right sequence. To tackle that, we propose a novel frameworknamed \textbf{C}ontrastive learning with \textbf{Bi}directional\textbf{T}ransformers for sequential recommendation (\textbf{CBiT}).Specifically, we first apply the slide window technique for long user sequencesin bidirectional Transformers, which allows for a more fine-grained division ofuser sequences. Then we combine the cloze task mask and the dropout mask togenerate high-quality positive samples and perform multi-pair contrastivelearning, which demonstrates better performance and adaptability compared withthe normal one-pair contrastive learning. Moreover, we introduce a noveldynamic loss reweighting strategy to balance between the cloze task loss andthe contrastive loss. Experiment results on three public benchmark datasetsshow that our model outperforms state-of-the-art models for sequentialrecommendation.

4. CCL4Rec：面向短视频推荐的对比学习上的对比框架

Title: CCL4Rec: Contrast over Contrastive Learning for Micro-video Recommendation

Published: 2022-08-17

Url: http://arxiv.org/abs/2208.08024v1

Authors: Shengyu Zhang,Bofang Li,Dong Yao,Fuli Feng,Jieming Zhu,Wenyan Fan,Zhou Zhao,Xiaofei He,Tat-seng Chua,Fei Wu

微视频推荐系统受到用户行为中普遍存在的噪声的影响，这可能会导致学习到的用户表示不加区分，并导致琐碎的推荐（例如热门项目）甚至是远远超出用户兴趣的奇怪推荐。对比学习是一种新兴技术，用于通过随机数据增强来学习区分表示。然而，由于忽略了用户行为中的噪声并平等地对待所有增强样本，现有的对比学习框架不足以学习区分用户表示的推荐。为了弥合这一研究差距，我们提出了用于训练推荐模型的 Contrast over Contrastive Learning 框架，名为 CCL4Rec，该框架通过进一步对比增强的正/负与自适应拉/推强度来模拟不同增强视图的细微差别，即对比"对比学习"。为了适应这些对比，我们设计了硬度感知增强，跟踪查询用户中被替换行为的重要性和替代品的相关性，从而确定增强的正/负的质量。硬度感知增强还允许可控对比学习，领先性能提升和稳健的培训。通过这种方式，CCL4Rec 捕获了给定用户历史行为的细微差别，从而明确地将学习到的用户表示与噪声行为的影响隔离开来。我们在两个微视频推荐基准上进行了广泛的实验，证明了模型参数少得多的 CCL4Rec 可以实现与现有最先进方法相当的性能，并将训练/推理速度提高几个数量级。

Micro-video recommender systems suffer from the ubiquitous noises in users'behaviors, which might render the learned user representation indiscriminating,and lead to trivial recommendations (e.g., popular items) or even weird onesthat are far beyond users' interests. Contrastive learning is an emergenttechnique for learning discriminating representations with random dataaugmentations. However, due to neglecting the noises in user behaviors andtreating all augmented samples equally, the existing contrastive learningframework is insufficient for learning discriminating user representations inrecommendation. To bridge this research gap, we propose the Contrast overContrastive Learning framework for training recommender models, named CCL4Rec,which models the nuances of different augmented views by further contrastingaugmented positives/negatives with adaptive pulling/pushing strengths, i.e.,the contrast over (vanilla) contrastive learning. To accommodate thesecontrasts, we devise the hardness-aware augmentations that track the importanceof behaviors being replaced in the query user and the relatedness ofsubstitutes, and thus determining the quality of augmented positives/negatives.The hardness-aware augmentation also permits controllable contrastive learning,leading to performance gains and robust training. In this way, CCL4Rec capturesthe nuances of historical behaviors for a given user, which explicitly shieldsoff the learned user representation from the effects of noisy behaviors. Weconduct extensive experiments on two micro-video recommendation benchmarks,which demonstrate that CCL4Rec with far less model parameters could achievecomparable performance to existing state-of-the-art method, and improve thetraining/inference speed by several orders of magnitude.

5. Re4：学习重新对比、重新参与、重新构建多兴趣推荐

Title: Re4: Learning to Re-contrast, Re-attend, Re-construct for Multi-interest Recommendation

Published: 2022-08-17

Url: http://arxiv.org/abs/2208.08011v1

Authors: Shengyu Zhang,Lingxiao Yang,Dong Yao,Yujie Lu,Fuli Feng,Zhou Zhao,Tat-seng Chua,Fei Wu

有效地代表用户是现代推荐系统的核心。由于用户的兴趣自然会表现出多个方面，因此开发多兴趣框架进行推荐，而不是用整体嵌入来代表每个用户，这一点越来越重要。尽管它们很有效，但现有方法仅利用编码器（前向流）来表示兴趣的多个方面。然而，如果没有明确的正则化，兴趣嵌入可能不会彼此区分，也不会在语义上反映具有代表性的历史项目。为此，我们提出了 Re4 框架，它利用反向流来重新检查每个兴趣嵌入。具体来说，Re4 封装了三个反向流，即 1）重新对比，通过对比学习使每个兴趣嵌入与其他兴趣区分开来；2) Re-attend，保证前向流中的interest-item相关性估计与最终推荐使用的准则一致；3) 重构，保证每个兴趣嵌入能够在语义上反映与相应兴趣相关的代表性项目的信息。我们在 ComiRec 上展示了新颖的前后多兴趣范式，并在三个真实世界的数据集上进行了广泛的实验。实证研究证实 Re4 有助于学习学习不同且有效的多兴趣表征。

Effectively representing users lie at the core of modern recommender systems.Since users' interests naturally exhibit multiple aspects, it is of increasinginterest to develop multi-interest frameworks for recommendation, rather thanrepresent each user with an overall embedding. Despite their effectiveness,existing methods solely exploit the encoder (the forward flow) to representmultiple aspects of interests. However, without explicit regularization, theinterest embeddings may not be distinct from each other nor semanticallyreflect representative historical items. Towards this end, we propose the Re4framework, which leverages the backward flow to reexamine each interestembedding. Specifically, Re4 encapsulates three backward flows, i.e., 1)Re-contrast, which drives each interest embedding to be distinct from otherinterests using contrastive learning; 2) Re-attend, which ensures theinterest-item correlation estimation in the forward flow to be consistent withthe criterion used in final recommendation; and 3) Re-construct, which ensuresthat each interest embedding can semantically reflect the information ofrepresentative items that relate to the corresponding interest. We demonstratethe novel forward-backward multi-interest paradigm on ComiRec, and performextensive experiments on three real-world datasets. Empirical studies validatethat Re4 helps to learn learning distinct and effective multi-interestrepresentations.

6. 具有对比反事实学习的干预性推荐，以更好地理解用户偏好

Title: Interventional Recommendation with Contrastive Counterfactual Learning for Better Understanding User Preferences

Published: 2022-08-13

Url: http://arxiv.org/abs/2208.06746v1

Authors: Guanglin Zhou,Chengkai Huang,Xiaocong Chen,Lina Yao,Xiwei Xu,Chen Wang,Liming Zhu

最近，人们对在因果推理的背景下制定建议的兴趣高涨。这些研究将推荐视为对因果推理的一种干预，并将用户的偏好设置为干预效果，以提高推荐系统的泛化能力。推荐系统因果推理领域的许多研究都集中在利用来自因果社区的倾向得分，以减少偏差，同时引起额外的方差。或者，一些研究表明存在一组来自随机对照试验的无偏数据，同时它需要满足某些在实践中可能具有挑战性的假设。在本文中，我们首先设计了一个因果图来表示推荐系统的数据生成和传播过程。然后，我们揭示了潜在的暴露机制使最大似然估计（MLE）偏向于观察反馈。为了从数据背后的因果关系中找出用户的偏好，我们利用后门调整和do-calculus，引入了介入推荐模型（IREC）。此外，考虑到混杂因素可能无法测量，我们提出了对比反事实用于模拟干预的学习方法（CCL）。此外，我们提出了两种额外的新抽样策略，并展示了一个有趣的发现，即从反事实集合中抽样有助于提高性能。我们对两个真实世界的数据集进行了广泛的实验，以评估和分析我们的模型 IREC-CCL 在无偏测试集上的性能。实验结果表明我们的模型优于最先进的方法。

Recently, there has been a surging interest in formulating recommendations inthe context of causal inference. The studies regard the recommendation as anintervention in causal inference and frame the users' preferences asinterventional effects to improve recommender systems' generalization. Manystudies in the field of causal inference for recommender systems have beenfocusing on utilizing propensity scores from the causal community that reducethe bias while inducing additional variance. Alternatively, some studiessuggest the existence of a set of unbiased data from randomized controlledtrials while it requires to satisfy certain assumptions that may be challengingin practice. In this paper, we first design a causal graph representingrecommender systems' data generation and propagation process. Then, we revealthat the underlying exposure mechanism biases the maximum likelihood estimation(MLE) on observational feedback. In order to figure out users' preferences interms of causality behind data, we leverage the back-door adjustment anddo-calculus, which induces an interventional recommendation model (IREC).Furthermore, considering the confounder may be inaccessible for measurement, wepropose a contrastive counterfactual learning method (CCL) for simulating theintervention. In addition, we present two extra novel sampling strategies andshow an intriguing finding that sampling from counterfactual sets contributesto superior performance. We perform extensive experiments on two real-worlddatasets to evaluate and analyze the performance of our model IREC-CCL onunbiased test sets. Experimental results demonstrate our model outperforms thestate-of-the-art methods.

7. 通过对比学习建模音乐推荐中负面偏好

Title: Exploiting Negative Preference in Content-based Music Recommendation with Contrastive Learning

Published: 2022-07-28

Url: http://arxiv.org/abs/2207.13909v1

Authors: Minju Park,Kyogu Lee

随着机器学习的发展，正在引入高级音乐推荐系统。然而，设计一个音乐推荐系统是必不可少的，它可以通过了解用户的音乐品味而不是模型的复杂性来提高用户满意度。尽管一些与利用负面偏好的音乐推荐系统相关的研究显示了性能改进，但缺乏对它们如何导致更好推荐的解释。在这项工作中，我们通过比较音乐推荐模型与对比学习利用偏好（CLEP）但具有三种不同的训练策略 - 同时利用积极和消极偏好（CLEP-PN）、仅积极偏好（CLEP）来分析消极偏好在用户音乐品味中的作用。-P) 和仅负数 (CLEP-N)。我们通过使用通过调查获得的少量个性化数据验证每个系统来评估负面偏好的有效性，并进一步阐明在音乐推荐中利用负面偏好的可能性。我们的实验结果表明，CLEP-N 在准确率和误报率方面优于其他两种。此外，无论不同类型的前端音乐特征提取器，所提出的训练策略都产生了一致的趋势，证明了所提出方法的稳定性。

Advanced music recommendation systems are being introduced along with thedevelopment of machine learning. However, it is essential to design a musicrecommendation system that can increase user satisfaction by understandingusers' music tastes, not by the complexity of models. Although several studiesrelated to music recommendation systems exploiting negative preferences haveshown performance improvements, there was a lack of explanation on how they ledto better recommendations. In this work, we analyze the role of negativepreference in users' music tastes by comparing music recommendation models withcontrastive learning exploiting preference (CLEP) but with three differenttraining strategies - exploiting preferences of both positive and negative(CLEP-PN), positive only (CLEP-P), and negative only (CLEP-N). We evaluate theeffectiveness of the negative preference by validating each system with a smallamount of personalized data obtained via survey and further illuminate thepossibility of exploiting negative preference in music recommendations. Ourexperimental results show that CLEP-N outperforms the other two in accuracy andfalse positive rate. Furthermore, the proposed training strategies produced aconsistent tendency regardless of different types of front-end musical featureextractors, proving the stability of the proposed method.

8. 时尚交互推荐系统中的对比学习

Title: Contrastive Learning for Interactive Recommendation in Fashion

Published: 2022-07-25

Url: http://arxiv.org/abs/2207.12033v1

Authors: Karin Sevegnani,Arjun Seshadri,Tian Wang,Anurag Beniwal,Julian McAuley,Alan Lu,Gerard Medioni

推荐系统和搜索对于促进在线时尚平台的个性化和易于浏览都是必不可少的。然而，这两种工具通常独立运作，无法将推荐系统的优势与搜索系统处理用户查询的能力相结合，以准确捕捉用户口味。我们通过基于用户提供的文本请求自动推荐个性化时尚物品，提出了一种新的解决方法。我们提出的模型 WhisperLite 使用对比学习从自然语言文本中捕获用户意图，并提高时尚产品的推荐质量。WhisperLite 将 CLIPembeddings 的强度与用于个性化的附加神经网络层相结合，并使用基于二元交叉熵和对比损失的复合损失函数进行训练。该模型在从在线零售时尚商店收集的真实数据集以及在不同电子商务领域（如餐厅、电影和电视节目、服装和鞋评。我们还进行了一项用户研究，捕捉用户对模型推荐项目相关性的判断，确认 WhisperLite 推荐在在线环境中的相关性。

Recommender systems and search are both indispensable in facilitatingpersonalization and ease of browsing in online fashion platforms. However, thetwo tools often operate independently, failing to combine the strengths ofrecommender systems to accurately capture user tastes with search systems'ability to process user queries. We propose a novel remedy to this problem byautomatically recommending personalized fashion items based on a user-providedtext request. Our proposed model, WhisperLite, uses contrastive learning tocapture user intent from natural language text and improves the recommendationquality of fashion products. WhisperLite combines the strength of CLIPembeddings with additional neural network layers for personalization, and istrained using a composite loss function based on binary cross entropy andcontrastive loss. The model demonstrates a significant improvement in offlinerecommendation retrieval metrics when tested on a real-world dataset collectedfrom an online retail fashion store, as well as widely used open-sourcedatasets in different e-commerce domains, such as restaurants, movies and TVshows, clothing and shoe reviews. We additionally conduct a user study thatcaptures user judgements on the relevance of the model's recommended items,confirming the relevancy of WhisperLite's recommendations in an online setting.

9. 预测和对比：推荐的双辅助学习

Title: Predictive and Contrastive: Dual-Auxiliary Learning for Recommendation

Published: 2022-07-23

Url: http://arxiv.org/abs/2203.03982v2

Authors: Yinghui Tao,Min Gao,Junliang Yu,Zongwei Wang,Qingyu Xiong,Xu Wang

自监督学习（SSL）最近在推荐方面取得了突出的成功。通过设置辅助任务（预测性或对比性），SSL 可以在无需人工注释的情况下从原始数据中发现监督信号，从而极大地缓解用户-项目交互稀疏的问题。然而，大多数基于 SSL 的推荐模型依赖于通用的辅助任务，例如，最大化从原始交互图和扰动交互图中学习的节点表示之间的对应关系，这与推荐任务明确无关。因此，基于推荐数据的异构图所反映的社会关系和项目类别所反映的丰富语义没有得到充分利用。为了探索特定于推荐的辅助任务，我们首先对异构交互数据进行定量分析，并发现交互之间存在很强的正相关关系。以及由元路径引起的用户项路径的数量。基于这一发现，我们设计了两个与目标任务紧密耦合的辅助任务（一个是预测性的，另一个是对比性的），以将推荐与隐藏在正相关中的自我监督信号联系起来。最后，开发了一个模型不可知的双辅助学习（DUAL）框架，它统一了 SSL 和推荐任务。在三个真实世界数据集上进行的大量实验表明，DUAL 可以显着提高推荐度，达到最先进的性能。

Self-supervised learning (SSL) recently has achieved outstanding success onrecommendation. By setting up an auxiliary task (either predictive orcontrastive), SSL can discover supervisory signals from the raw data withouthuman annotation, which greatly mitigates the problem of sparse user-iteminteractions. However, most SSL-based recommendation models rely ongeneral-purpose auxiliary tasks, e.g., maximizing correspondence between noderepresentations learned from the original and perturbed interaction graphs,which are explicitly irrelevant to the recommendation task. Accordingly, therich semantics reflected by social relationships and item categories, which liein the recommendation data-based heterogeneous graphs, are not fully exploited.To explore recommendation-specific auxiliary tasks, we first quantitativelyanalyze the heterogeneous interaction data and find a strong positivecorrelation between the interactions and the number of user-item paths inducedby meta-paths. Based on the finding, we design two auxiliary tasks that aretightly coupled with the target task (one is predictive and the other one iscontrastive) towards connecting recommendation with the self-supervisionsignals hiding in the positive correlation. Finally, a model-agnosticDUal-Auxiliary Learning (DUAL) framework which unifies the SSL andrecommendation tasks is developed. The extensive experiments conducted on threereal-world datasets demonstrate that DUAL can significantly improverecommendation, reaching the state-of-the-art performance.

10. 基于多粒度项目的对比推荐

Title: Multi-granularity Item-based Contrastive Recommendation

Published: 2022-07-04

Url: http://arxiv.org/abs/2207.01387v1

Authors: Ruobing Xie,Zhijie Qiu,Bo Zhang,Leyu Lin

对比学习（CL）在推荐方面显示了它的力量。然而，大多数基于 CL 的推荐模型构建它们的 CL 任务仅仅关注用户的方面，而忽略了项目中丰富多样的信息。在这项工作中，我们提出了一种新颖的基于多粒度项目的对比学习（MicRec）框架，用于推荐的匹配阶段（即候选生成），它系统地将多方面的项目相关信息引入使用 CL 的表示学习。具体来说，我们将基于三项的 CL 任务构建为一组即插即用的辅助目标，以捕获特征、语义和会话级别的项相关性。特征级项目 CL 旨在通过项目及其增强来学习细粒度的特征级项目相关性。语义级项目 CL 关注语义相关项目之间的粗粒度语义相关性。会话级项目 CL 从用户在所有会话中的顺序行为中突出项目的全局行为相关性。在实验中，我们对真实世界的数据集进行离线和在线评估，在这些数据集上，MicRecachieves 比竞争基线有显着改进。此外，我们进一步验证了三个 CL 任务的有效性以及 MicRec 在不同匹配模型上的普遍性。所提出的 MicRec 有效、高效、通用且易于部署，已部署在现实世界推荐系统上，影响数百万用户。源代码将在未来发布。

Contrastive learning (CL) has shown its power in recommendation. However,most CL-based recommendation models build their CL tasks merely focusing on theuser's aspects, ignoring the rich diverse information in items. In this work,we propose a novel Multi-granularity item-based contrastive learning (MicRec)framework for the matching stage (i.e., candidate generation) inrecommendation, which systematically introduces multi-aspect item-relatedinformation to representation learning with CL. Specifically, we build threeitem-based CL tasks as a set of plug-and-play auxiliary objectives to captureitem correlations in feature, semantic and session levels. The feature-levelitem CL aims to learn the fine-grained feature-level item correlations viaitems and their augmentations. The semantic-level item CL focuses on thecoarse-grained semantic correlations between semantically related items. Thesession-level item CL highlights the global behavioral correlations of itemsfrom users' sequential behaviors in all sessions. In experiments, we conductboth offline and online evaluations on real-world datasets, where MicRecachieves significant improvements over competitive baselines. Moreover, wefurther verify the effectiveness of three CL tasks as well as the universalityof MicRec on different matching models. The proposed MicRec is effective,efficient, universal, and easy to deploy, which has been deployed on areal-world recommendation system, affecting millions of users. The source codewill be released in the future.