会员服务 ·

基于图的推荐系统进展总结, 共包含11篇文献

2022 年 2 月 26 日 机器学习与推荐算法

收录2022.2最新基于Graph推荐的前沿研究工作。

1. 用于社交推荐的动态和静态表示的图神经网络

Title: Graph Neural Networks with Dynamic and Static Representations for Social Recommendation

Published: 2022-01-31

Url: http://arxiv.org/abs/2201.10751v2

Authors: Junfa Lin,Siyuan Chen,Jiahai Wang

基于图神经网络的推荐系统因其学习包括社交网络在内的各种边信息的优异能力而受到越来越多的研究兴趣。然而，以前的工作通常集中在建模用户上，对物品没有太多关注。此外，随着时间的推移，物品的吸引力可能会发生变化，这就像用户的动态兴趣一样，很少被考虑，物品之间的相关性也很少被考虑。为了克服这些局限性，本文提出了具有动态和静态表示的社交推荐图神经网络（GNN-DSR），它考虑了用户和物品的动态和静态表示，并考虑了它们的关系影响。GNN-DSR分别对用户兴趣和物品吸引力的短期动态和长期静态交互表示进行建模。此外，注意机制用于聚合用户对目标用户的社交影响以及相关物品对给定物品的影响。最后结合用户和物品的潜在因素进行预测。在三个真实世界推荐系统数据集上的实验验证了GNN-DSR的有效性。

Recommender systems based on graph neural networks receive increasing research interest due to their excellent ability to learn a variety of side information including social networks. However, previous works usually focus on modeling users, not much attention is paid to items. Moreover, the possible changes in the attraction of items over time, which is like the dynamic interest of users are rarely considered, and neither do the correlations among items. To overcome these limitations, this paper proposes graph neural networkswith dynamic and static representations for social recommendation (GNN-DSR), which considers both dynamic and static representations of users and items and incorporates their relational influence. GNN-DSR models the short-term dynamic and long-term static interactional representations of the user's interest andthe item's attraction, respectively. Furthermore, the attention mechanism isused to aggregate the social influence of users on the target user and thecorrelative items' influence on a given item. The final latent factors of user and item are combined to make a prediction. Experiments on three real-worldrecommender system datasets validate the effectiveness of GNN-DSR.

2. FairMod：通过图修改的公平链接预测和推荐

Title: FairMod: Fair Link Prediction and Recommendation via Graph Modification

Published: 2022-01-27

Url: http://arxiv.org/abs/2201.11596v1

Authors: Sean Current,Yuntian He,Saket Gurukar,Srinivasan Parthasarathy

随着机器学习在各个领域得到越来越广泛的应用，研究人员和ML工程师必须考虑模型可能带来的数据固有偏差。最近，许多研究表明，如果输入图有偏差，这种偏差也会被图神经网络（GNN）模型吸收。在这项工作中，我们的目标是通过修改输入图来减轻GNNS学习到的偏差。为此，我们提出了FairMod，这是一种FairGraph修改方法，包含三个公式：全局公平优化（GFO）、社区公平优化（CFO）和公平边缘加权（少数）模型。我们提出的模型在训练GNN的同时对输入图进行微观或宏观编辑，并学习在链接建议的背景下准确且公平的节点嵌入。我们在四个真实数据集上证明了我们的方法的有效性，并表明我们可以通过几个因素来提高推荐公平性，而链接预测精度的成本可以忽略不计。

As machine learning becomes more widely adopted across domains, it iscritical that researchers and ML engineers think about the inherent biases inthe data that may be perpetuated by the model. Recently, many studies haveshown that such biases are also imbibed in Graph Neural Network (GNN) models ifthe input graph is biased. In this work, we aim to mitigate the bias learned byGNNs through modifying the input graph. To that end, we propose FairMod, a FairGraph Modification methodology with three formulations: the Global FairnessOptimization (GFO), Community Fairness Optimization (CFO), and Fair EdgeWeighting (FEW) models. Our proposed models perform either microscopic ormacroscopic edits to the input graph while training GNNs and learn nodeembeddings that are both accurate and fair under the context of linkrecommendations. We demonstrate the effectiveness of our approach on four realworld datasets and show that we can improve the recommendation fairness byseveral factors at negligible cost to link prediction accuracy.

3. 基于会话的有效可解释推荐的因果关系和相关图建模

Title: Causality and Correlation Graph Modeling for Effective and Explainable Session-based Recommendation

Published: 2022-01-27

Url: http://arxiv.org/abs/2201.10782v2

Authors: Cong Geng,Huizi Wu,Hui Fang

最近，基于会话的推荐已经引起了人们的兴趣，它主要是基于一个非即时会话来预测用户下一个感兴趣的项目。大多数现有研究采用复杂的深度学习技术（例如，图形神经网络）进行有效的基于会话的推荐。然而，它们仅仅解决了项目之间的共现问题，却未能很好地区分因果关系和相关关系。考虑到项目之间因果关系和相关关系的不同解释和特点，在本研究中，我们提出了一种新的方法，称为CGSR，通过联合建模项目之间的因果关系和相关关系。特别是，我们通过同时考虑虚假因果关系问题，从会话中构建因果图和相关图。我们进一步设计了基于agraph神经网络的会话推荐方法。在三个数据集上的扩展实验表明，我们的模型在推荐准确性方面优于其他最先进的方法。此外，我们还提出了一个可解释的CGSR框架，并通过对Amazon数据集的案例研究证明了模型的可解释性。

Session-based recommendation which has been witnessed a booming interestrecently, focuses on predicting a user's next interested item(s) based on ananonymous session. Most existing studies adopt complex deep learning techniques(e.g., graph neural networks) for effective session-based recommendation.However, they merely address co-occurrence between items, but fail to welldistinguish causality and correlation relationship. Considering the variedinterpretations and characteristics of causality and correlation relationshipbetween items, in this study, we propose a novel method denoted as CGSR byjointly modeling causality and correlation relationship between items. Inparticular, we construct cause, effect and correlation graphs from sessions bysimultaneously considering the false causality problem. We further design agraph neural network-based method for session-based recommendation. Extensiveexperiments on three datasets show that our model outperforms otherstate-of-the-art methods in terms of recommendation accuracy. Moreover, wefurther propose an explainable framework on CGSR, and demonstrate theexplainability of our model via case studies on Amazon dataset.

4. 基于知识图的波形推荐：一种新的通信波形设计范式

Title: Knowledge Graph Based Waveform Recommendation: A New Communication Waveform Design Paradigm

Published: 2022-01-24

Url: http://arxiv.org/abs/2202.01926v1

Authors: Wei Huang,Tianfu Qi,Yundi Guan,Qihang Peng,Jun Wang

传统上，通信波形是由专家根据通信理论和他们的经验逐个设计的，这通常很费时费力。在本文中，我们从一个新的角度研究了波形设计，并提出了一种基于知识图（KG）的智能推荐系统的新波形设计范式。提出的范例旨在通过对现有波形进行结构化表征和表示，并智能地利用从中学习到的知识，提高设计效率。为了实现这一目标，我们首先构建了一个具有一阶邻域节点的通信波形知识图（CWKG），通过表示学习将波形的结构化语义知识和数值参数结合起来。在此基础上，我们进一步提出了一个智能通信波形推荐系统（CWRS）来生成候选波形。在CWRS中，根据基于KG的波形表示特征提取的特点，引入了一种改进的与信道无关、与空间相关的进化1D算子，并采用多头自注意加权各分量的影响进行特征融合。同时，采用基于多层感知器的协同过滤来评估需求与候选波形之间的匹配程度。仿真结果表明，基于CWKG的CWR能够自动推荐高可靠性的波形指标。

Traditionally, a communication waveform is designed by experts based oncommunication theory and their experiences on a case-by-case basis, which isusually laborious and time-consuming. In this paper, we investigate thewaveform design from a novel perspective and propose a new waveform designparadigm with the knowledge graph (KG)-based intelligent recommendation system.The proposed paradigm aims to improve the design efficiency by structuralcharacterization and representations of existing waveforms and intelligentlyutilizing the knowledge learned from them. To achieve this goal, we first builda communication waveform knowledge graph (CWKG) with a first-order neighbornode, for which both structured semantic knowledge and numerical parameters ofa waveform are integrated by representation learning. Based on the developedCWKG, we further propose an intelligent communication waveform recommendationsystem (CWRS) to generate waveform candidates. In the CWRS, an improvedinvolution1D operator, which is channel-agnostic and space-specific, isintroduced according to the characteristics of KG-based waveform representationfor feature extraction, and the multi-head self-attention is adopted to weighthe influence of various components for feature fusion. Meanwhile, multilayerperceptron-based collaborative filtering is used to evaluate the matchingdegree between the requirement and the waveform candidate. Simulation resultsshow that the proposed CWKG-based CWRS can automatically recommend waveformcandidates with high reliability.

5. 基于邻近图的强化路由算法

Title: Reinforcement Routing on Proximity Graph for Efficient Recommendation

Published: 2022-01-23

Url: http://arxiv.org/abs/2201.09290v1

Authors: Chao Feng,Defu Lian,Xiting Wang,Zheng liu,Xing Xie,Enhong Chen

我们关注最大内积搜索（MIPS），这是许多机器学习社区的一个基本问题。在给定的查询中，MIPS会查找具有最大内积的最相似项。由于内积是一个非度量函数，通常定义在度量空间上的最近邻搜索（NNS）方法对MIPS问题并没有表现出令人满意的性能。然而，与度量函数相比，内积表现出许多良好的性质，例如避免消失和爆炸梯度。因此，内积在许多推荐系统中得到了广泛的应用，这使得高效的最大内积搜索成为加速许多推荐系统的关键。基于图的NNS问题求解方法显示了与其他类方法相比的优越性。数据库的每个数据点都映射到近似图的一个节点。数据库中的最近邻搜索可以转换为邻近图上的路由，以查找查询的最近邻。该技术可用于解决MIPS问题。我们在邻近图上用query搜索具有最大内积的项，而不是搜索查询的最近邻居。在本文中，我们提出了一个强化模型，当我们缺乏训练查询的基本事实时，训练代理自动在接近图上搜索MIPSP问题。如果我们知道一些训练查询的基本事实，我们的模型也可以通过模仿学习来利用这些基本事实来提高代理的搜索能力。通过实验，我们可以看到，我们提出的强化学习与模仿学习相结合的模式显示出了与最先进的方法相比的优越性

We focus on Maximum Inner Product Search (MIPS), which is an essentialproblem in many machine learning communities. Given a query, MIPS finds themost similar items with the maximum inner products. Methods for NearestNeighbor Search (NNS) which is usually defined on metric space don't exhibitthe satisfactory performance for MIPS problem since inner product is anon-metric function. However, inner products exhibit many good propertiescompared with metric functions, such as avoiding vanishing and explodinggradients. As a result, inner product is widely used in many recommendationsystems, which makes efficient Maximum Inner Product Search a key for speedingup many recommendation systems. Graph based methods for NNS problem show the superiorities compared withother class methods. Each data point of the database is mapped to a node of theproximity graph. Nearest neighbor search in the database can be converted toroute on the proximity graph to find the nearest neighbor for the query. Thistechnique can be used to solve MIPS problem. Instead of searching the nearestneighbor for the query, we search the item with maximum inner product withquery on the proximity graph. In this paper, we propose a reinforcement modelto train an agent to search on the proximity graph automatically for MIPSproblem if we lack the ground truths of training queries. If we know the groundtruths of some training queries, our model can also utilize these ground truthsby imitation learning to improve the agent's search ability. By experiments, wecan see that our proposed mode which combines reinforcement learning withimitation learning shows the superiorities over the state-of-the-art methods

6. 为第三方图书馆提供升级计划：使用迁移图的推荐系统

Title: Providing Upgrade Plans for Third-party Libraries: A Recommender System using Migration Graphs

Published: 2022-01-20

Url: http://arxiv.org/abs/2201.08201v1

Authors: Riccardo Rubei,Davide Di Ruscio,Claudio Di Sipio,Juri Di Rocco,Phuong T. Nguyen

在软件项目的开发过程中，开发人员经常需要升级第三方库（TPL），以使他们的代码与所用库提供的最新功能保持同步。在大多数情况下，升级使用过的TPL是一项复杂且容易出错的活动，必须谨慎执行，以限制对依赖于正在升级的库的软件项目的连锁反应。在本文中，我们提出EvoPlan作为一种级别的方法，在给定一对库版本作为输入的情况下，推荐不同的升级计划。特别是，在将当前库版本升级到所需更新版本可能遵循的不同路径中，EvoPlan能够建议一个计划，该计划可以潜在地最小化将客户机代码从库的当前版本迁移到目标版本所需的工作。使用信息检索中使用的常规指标，即精确度、召回率和F-度量，在一个精心策划的数据集上对该方法进行了评估。实验结果表明，考虑到计划规范中的两个不同标准，即迁移路径的流行程度和GitHub中已遵循推荐迁移路径的项目的开放和关闭问题的数量，EvoPlan获得了令人鼓舞的预测性能。

During the development of a software project, developers often need toupgrade third-party libraries (TPLs), aiming to keep their code up-to-date withthe newest functionalities offered by the used libraries. In most cases,upgrading used TPLs is a complex and error-prone activity that must becarefully carried out to limit the ripple effects on the software project thatdepends on the libraries being upgraded. In this paper, we propose EvoPlan as anovel approach to the recommendation of different upgrade plans given a pair oflibrary-version as input. In particular, among the different paths that can bepossibly followed to upgrade the current library version to the desired updatedone, EvoPlan is able to suggest the plan that can potentially minimize theefforts being needed to migrate the code of the clients from the library'scurrent release to the target one. The approach has been evaluated on a curateddataset using conventional metrics used in Information Retrieval, i.e.,precision, recall, and F-measure. The experimental results show that EvoPlanobtains an encouraging prediction performance considering two differentcriteria in the plan specification, i.e., the popularity of migration paths andthe number of open and closed issues in GitHub for those projects that havealready followed the recommended migration paths.

7. 用于序列推荐的位置增强和时间感知图卷积网络

Title: Position-enhanced and Time-aware Graph Convolutional Network for Sequential Recommendations

Published: 2022-01-14

Url: http://arxiv.org/abs/2107.05235v2

Authors: Liwei Huang,Yutao Ma,Yanbo Liu,Bohong,Du,Shuliang Wang,Deyi Li

现有的基于深度学习的顺序推荐方法大多采用递归神经网络结构或自我注意来建模用户历史行为之间的顺序模式和时间影响，并在特定时间学习用户的偏好。然而，这些方法有两个主要缺点。首先，他们专注于从以用户为中心的角度对用户的动态状态进行建模，而总是忽略项目的动态。第二，大多数用户只处理一阶用户项目交互，并且不考虑用户和项目之间的高阶连通性，最近被证明有助于序列推荐。为了解决上述问题，在本文中，我们尝试用二部图结构来建模用户项交互，并提出了一种新的基于位置增强和时间感知的图形进化网络（PTGCN）的顺序推荐方法。PTGCN通过定义位置增强和时间感知的图卷积运算，并使用自我注意聚合器在二部图上同时学习用户和项目的动态表示，对用户-项目交互之间的顺序模式和时间动态进行建模。同时，通过多层图形的叠加，实现了用户与物品之间的高阶连通。为了证明PTGCN的有效性，我们在三个不同大小的真实数据集上对PTGCN进行了综合评估，并与一些竞争基线进行了比较。实验结果表明，PTGCN在两种常用的排名评估指标方面优于几种最先进的模型。

Most of the existing deep learning-based sequential recommendation approachesutilize the recurrent neural network architecture or self-attention to modelthe sequential patterns and temporal influence among a user's historicalbehavior and learn the user's preference at a specific time. However, thesemethods have two main drawbacks. First, they focus on modeling users' dynamicstates from a user-centric perspective and always neglect the dynamics of itemsover time. Second, most of them deal with only the first-order user-iteminteractions and do not consider the high-order connectivity between users anditems, which has recently been proved helpful for the sequentialrecommendation. To address the above problems, in this article, we attempt tomodel user-item interactions by a bipartite graph structure and propose a newrecommendation approach based on a Position-enhanced and Time-aware GraphConvolutional Network (PTGCN) for the sequential recommendation. PTGCN modelsthe sequential patterns and temporal dynamics between user-item interactions bydefining a position-enhanced and time-aware graph convolution operation andlearning the dynamic representations of users and items simultaneously on thebipartite graph with a self-attention aggregator. Also, it realizes thehigh-order connectivity between users and items by stacking multi-layer graphconvolutions. To demonstrate the effectiveness of PTGCN, we carried out acomprehensive evaluation of PTGCN on three real-world datasets of differentsizes compared with a few competitive baselines. Experimental results indicatethat PTGCN outperforms several state-of-the-art models in terms of twocommonly-used evaluation metrics for ranking.

8. 基于会话推荐的非纠缠图神经网络

Title: Disentangled Graph Neural Networks for Session-based Recommendation

Published: 2022-01-11

Url: http://arxiv.org/abs/2201.03482v2

Authors: Ansong Li,Zhiyong Cheng,Fan Liu,Zan Gao,Weili Guan,Yuxin Peng

基于会话的推荐（Session-based Recommension，SBR）由于仅利用当前会话中有限的用户行为历史而具有巨大的实用价值，近年来受到了越来越多的研究关注。现有的方法通常是在项目层面上学习会话嵌入，即在有或没有为项目分配注意权重的情况下，聚合项目的嵌入。然而，他们忽略了一个事实，即用户采用某个项目的意图是由该项目的某些因素驱动的（例如，一部电影的主要演员）。换句话说，他们没有在因子级别探索用户更细粒度的兴趣，以生成会话嵌入，从而导致次优性能。为了解决这个问题，我们提出了一种新的方法，称为解纠缠图神经网络（Disen-GNN），在考虑每个项目的因素水平注意的情况下，捕获会话目的。具体来说，我们首先使用非纠缠学习技术将项嵌入转换为多个因子的嵌入，然后使用门控图形神经网络（GGNN）根据为每个因子计算的项邻接相似矩阵明智地学习嵌入因子。此外，采用距离相关性来增强每对因素之间的独立性。在用独立的因素表示每个项目后，设计了一个注意机制，以了解用户对会话中每个项目的不同因素的意图。会话嵌入是通过将项目嵌入与每个项目因素的注意权重聚合而生成的。为此，我们的模型在因素层面上考虑用户意图，以推断会话中的用户目的。在三个基准数据集上的大量实验表明，我们的方法优于现有方法。

Session-based recommendation (SBR) has drawn increasingly research attentionin recent years, due to its great practical value by only exploiting thelimited user behavior history in the current session. Existing methodstypically learn the session embedding at the item level, namely, aggregatingthe embeddings of items with or without the attention weights assigned toitems. However, they ignore the fact that a user's intent on adopting an itemis driven by certain factors of the item (e.g., the leading actors of anmovie). In other words, they have not explored finer-granularity interests ofusers at the factor level to generate the session embedding, leading tosub-optimal performance. To address the problem, we propose a novel methodcalled Disentangled Graph Neural Network (Disen-GNN) to capture the sessionpurpose with the consideration of factor-level attention on each item.Specifically, we first employ the disentangled learning technique to cast itemembeddings into the embedding of multiple factors, and then use the gated graphneural network (GGNN) to learn the embedding factor-wisely based on the itemadjacent similarity matrix computed for each factor. Moreover, the distancecorrelation is adopted to enhance the independence between each pair offactors. After representing each item with independent factors, an attentionmechanism is designed to learn user intent to different factors of each item inthe session. The session embedding is then generated by aggregating the itemembeddings with attention weights of each item's factors. To this end, ourmodel takes user intents at the factor level into account to infer the userpurpose in a session. Extensive experiments on three benchmark datasetsdemonstrate the superiority of our method over existing methods.

9. 基于图的注意推荐

Title: Attention-Based Recommendation On Graphs

Published: 2022-01-04

Url: http://arxiv.org/abs/2201.05499v1

Authors: Taher Hekmatfar,Saman Haratizadeh,Parsa Razban,Sama Goliaei

图形神经网络（GNN）在不同的任务中表现出了显著的性能。然而，关于推荐系统的GNN研究还很少。GCNas是一种GNN，可以为图形中的不同实体提取高质量的嵌入。在协同过滤任务中，核心问题是找出实体在预测目标用户未来行为方面的信息量。使用注意机制，当底层数据被建模为图形时，我们可以让GCN进行这样的分析。在这项研究中，我们提出了一个基于模型的推荐系统，该系统在推荐图上应用注意机制和空间GCN来提取用户和项目的嵌入。注意机制告诉GCN相关用户或项目对目标实体最终表示的影响程度。我们比较了GARec和一些基线算法在RMSE方面的性能。在不同的电影数据集中，该方法优于现有的基于模型的非图神经网络和图神经网络。

Graph Neural Networks (GNN) have shown remarkable performance in differenttasks. However, there are a few studies about GNN on recommender systems. GCNas a type of GNNs can extract high-quality embeddings for different entities ina graph. In a collaborative filtering task, the core problem is to find out howinformative an entity would be for predicting the future behavior of a targetuser. Using an attention mechanism, we can enable GCNs to do such an analysiswhen the underlying data is modeled as a graph. In this study, we proposedGARec as a model-based recommender system that applies an attention mechanismalong with a spatial GCN on a recommender graph to extract embeddings for usersand items. The attention mechanism tells GCN how much a related user or itemshould affect the final representation of the target entity. We compared theperformance of GARec against some baseline algorithms in terms of RMSE. Thepresented method outperforms existing model-based, non-graph neural networksand graph neural networks in different MovieLens datasets.

10. 具有协作指导的注意力知识感知图卷积网络个性化推荐

Title: Attentive Knowledge-aware Graph Convolutional Networks with Collaborative Guidance for Personalized Recommendation

Published: 2022-01-02

Url: http://arxiv.org/abs/2109.02046v2

Authors: Yankai Chen,Yaming Yang,Yujing Wang,Jing Bai,Xiangchen Song,Irwin King

为了缓解传统推荐系统（RSs）的数据稀疏性和冷启动问题，结合知识图（KG）补充辅助信息近年来引起了广泛关注。然而，在当前基于KG的RS模型中简单地集成KG并不一定能保证提高推荐性能，甚至可能削弱整体模型的能力。这是因为这些KG的构建独立于历史用户项交互的收集；因此，KGs中的信息可能并不总是有助于向所有用户推荐。在本文中，我们提出了一种基于专注知识的图卷积网络（CG-KGR），该网络具有个性化推荐的协作指导功能。CG-KGR是一种新的知识感知推荐模型，通过我们提出的协作指导机制，可以充分、连贯地学习KG和用户项目交互。具体来说，CG-KGR首先将历史交互封装为交互式信息摘要。然后CG KG将其作为指导，从KG中提取信息，最终提供更精确的个性化推荐。我们在两个推荐任务（即Top-K推荐和点击率（CTR）预测）上对四个真实数据集进行了扩展实验。实验结果表明，CG-KGR模型在顶级推荐的召回指标方面显著优于最新的模型1.4-27.0%。

To alleviate data sparsity and cold-start problems of traditional recommendersystems (RSs), incorporating knowledge graphs (KGs) to supplement auxiliaryinformation has attracted considerable attention recently. However, simplyintegrating KGs in current KG-based RS models is not necessarily a guarantee toimprove the recommendation performance, which may even weaken the holisticmodel capability. This is because the construction of these KGs is independentof the collection of historical user-item interactions; hence, information inthese KGs may not always be helpful for recommendation to all users. In this paper, we propose attentive Knowledge-aware Graph convolutionalnetworks with Collaborative Guidance for personalized Recommendation (CG-KGR).CG-KGR is a novel knowledge-aware recommendation model that enables ample andcoherent learning of KGs and user-item interactions, via our proposedCollaborative Guidance Mechanism. Specifically, CG-KGR first encapsulateshistorical interactions to interactive information summarization. Then CG-KGRutilizes it as guidance to extract information out of KGs, which eventuallyprovides more precise personalized recommendation. We conduct extensiveexperiments on four real-world datasets over two recommendation tasks, i.e.,Top-K recommendation and Click-Through rate (CTR) prediction. The experimentalresults show that the CG-KGR model significantly outperforms recentstate-of-the-art models by 1.4-27.0% in terms of Recall metric on Top-Krecommendation.

11. 基于双曲线几何的无标度图知识感知推荐模型

Title: Modeling Scale-free Graphs with Hyperbolic Geometry for Knowledge-aware Recommendation

Published: 2022-01-02

Url: http://arxiv.org/abs/2108.06468v3

Authors: Yankai Chen,Menglin Yang,Yingxue Zhang,Mengchen Zhao,Ziqiao Meng,Jianye Hao,Irwin King

为了缓解传统推荐系统的数据稀疏性和冷启动问题，将知识图（KG）与补充辅助信息相结合近来受到了广泛关注。通过将KG与用户项目交互统一成一个三元图，最近的工作探索了图拓扑，以学习用户和项目的低维表示，并具有丰富的语义。然而，这些现实世界中的三部图通常是无标度的，其固有的层次图结构在现有的研究中没有得到足够的重视，因此导致了次优的推荐性能。为了解决这个问题并提供更准确的推荐，我们提出了一种基于双曲线几何的知识感知推荐方法，即洛伦兹知识增强图卷积网络推荐方法（LKGR）。LKGR有助于在数据统一后更好地建模无标度三元图。具体地说，我们在双曲空间中使用不同的信息传播策略来明确地编码来自历史交互和KG的异质信息。我们提出的知识感知机制使模型能够自动测量信息贡献，在双曲线空间中产生一致的信息聚合。在三个真实基准上进行的大量实验表明，LKGR的性能比最先进的方法高出3.6-15.3%Recall@20最重要的是表扬。

Aiming to alleviate data sparsity and cold-start problems of traditionalrecommender systems, incorporating knowledge graphs (KGs) to supplementauxiliary information has recently gained considerable attention. Via unifyingthe KG with user-item interactions into a tripartite graph, recent worksexplore the graph topologies to learn the low-dimensional representations ofusers and items with rich semantics. However, these real-world tripartitegraphs are usually scale-free, the intrinsic hierarchical graph structures ofwhich are underemphasized in existing works, consequently, leading tosuboptimal recommendation performance. To address this issue and provide more accurate recommendation, we propose aknowledge-aware recommendation method with the hyperbolic geometry, namelyLorentzian Knowledge-enhanced Graph convolutional networks for Recommendation(LKGR). LKGR facilitates better modeling of scale-free tripartite graphs afterthe data unification. Specifically, we employ different information propagationstrategies in the hyperbolic space to explicitly encode heterogeneousinformation from historical interactions and KGs. Our proposed knowledge-awareattention mechanism enables the model to automatically measure the informationcontribution, producing the coherent information aggregation in the hyperbolicspace. Extensive experiments on three real-world benchmarks demonstrate thatLKGR outperforms state-of-the-art methods by 3.6-15.3% of Recall@20 on Top-Krecommendation.