会员服务 ·

跨域推荐最新前沿工作进展汇总

2022 年 9 月 29 日 机器学习与推荐算法

嘿，记得给“机器学习与推荐算法”添加星标

收录最新跨域推荐的前沿研究工作。

1. 用于跨领域推荐的知识感知神经集合矩阵分解

Title: Knowledge-aware Neural Collective Matrix Factorization for Cross-domain Recommendation

Published: 2022-06-27

Url: http://arxiv.org/abs/2206.13255v1

Authors: Li Zhang,Yan Ge,Jun Ma,Jianmo Ni,Haiping Lu

跨域推荐（CDR）可以帮助客户在不同领域找到更令人满意的项目。现有的CDR模型主要使用普通用户或映射函数作为域之间的桥梁，但在跨域充分利用额外知识方面的探索非常有限。在本文中，我们提出将知识图（KG）合并到CDR中，使不同主题中的项目能够共享知识。为此，我们首先从Freebase KG和亚马逊评论数据的子集（两个域对：电影音乐、电影书籍）构建一个新的数据集mazonkg4cdr。这个新的数据集有助于将知识链接到CDR的域内和跨域项目。然后，我们提出了一个新的框架，KG感知神经集体矩阵分解（KG NeuCMF），利用KG丰富项目表示。它首先通过图卷积自动编码器学习项目嵌入，以从KG中的相邻和高阶邻域中获取领域特定和领域一般知识。然后，我们最大化从KG和用户项目矩阵中学习到的项目嵌入之间的互信息，以建立跨域关系，从而实现更好的CDR。最后，我们在新构建的数据集上进行了大量实验，并证明我们的模型显著优于性能最佳的基线。

Cross-domain recommendation (CDR) can help customers find more satisfyingitems in different domains. Existing CDR models mainly use common users ormapping functions as bridges between domains but have very limited explorationin fully utilizing extra knowledge across domains. In this paper, we propose toincorporate the knowledge graph (KG) for CDR, which enables items in differentdomains to share knowledge. To this end, we first construct a new datasetAmazonKG4CDR from the Freebase KG and a subset (two domain pairs: movies-music,movie-book) of Amazon Review Data. This new dataset facilitates linkingknowledge to bridge within- and cross-domain items for CDR. Then we propose anew framework, KG-aware Neural Collective Matrix Factorization (KG-NeuCMF),leveraging KG to enrich item representations. It first learns item embeddingsby graph convolutional autoencoder to capture both domain-specific anddomain-general knowledge from adjacent and higher-order neighbours in the KG.Then, we maximize the mutual information between item embeddings learned fromthe KG and user-item matrix to establish cross-domain relationships for betterCDR. Finally, we conduct extensive experiments on the newly constructed datasetand demonstrate that our model significantly outperforms the best-performingbaselines.

2. 召回中的跨域对比推荐

Title: Contrastive Cross-domain Recommendation in Matching

Published: 2022-06-21

Url: http://arxiv.org/abs/2112.00999v2

Authors: Ruobing Xie,Qi Liu,Liangdong Wang,Shukai Liu,Bo Zhang,Leyu Lin

跨域推荐（CDR）旨在借助于源域在目标域中提供更好的推荐结果，这在现实系统中得到了广泛的应用和探索。然而，匹配（即候选生成）模块中的CDR在表示学习和知识转移中都面临着数据稀疏性和流行度偏差问题。在这项工作中，我们提出了一种新的对比跨域推荐（CCDR）框架，用于匹配CDR。具体来说，我们构建了一个巨大的多元化参考网络来捕获反映用户不同兴趣的多个信息，并设计了一个域内对比学习（intra-CL）和三个域间对比学习（inter-CL）任务，以更好地进行表征学习和知识转移。内部CL通过图形增强在目标域内实现更有效和平衡的训练，而内部CL从用户、分类和邻居方面构建不同类型的跨域交互。在实验中，CCDR在现实系统中的离线和在线评估方面都取得了显著的改进。目前，我们已经在微信头条新闻上部署了我们的CCDR，影响了大量用户。源代码位于https://github.com/lqfarmer/CCDR.

Cross-domain recommendation (CDR) aims to provide better recommendationresults in the target domain with the help of the source domain, which iswidely used and explored in real-world systems. However, CDR in the matching(i.e., candidate generation) module struggles with the data sparsity andpopularity bias issues in both representation learning and knowledge transfer.In this work, we propose a novel Contrastive Cross-Domain Recommendation (CCDR)framework for CDR in matching. Specifically, we build a huge diversifiedpreference network to capture multiple information reflecting user diverseinterests, and design an intra-domain contrastive learning (intra-CL) and threeinter-domain contrastive learning (inter-CL) tasks for better representationlearning and knowledge transfer. The intra-CL enables more effective andbalanced training inside the target domain via a graph augmentation, while theinter-CL builds different types of cross-domain interactions from user,taxonomy, and neighbor aspects. In experiments, CCDR achieves significantimprovements on both offline and online evaluations in a real-world system.Currently, we have deployed our CCDR on WeChat Top Stories, affecting plenty ofusers. The source code is in https://github.com/lqfarmer/CCDR.

3. 强化学习增强共享帐户跨域顺序推荐

Title: Reinforcement Learning-enhanced Shared-account Cross-domain Sequential Recommendation

Published: 2022-06-16

Url: http://arxiv.org/abs/2206.08088v1

Authors: Lei Guo,Jinyu Zhang,Tong Chen,Xinhua Wang,Hongzhi Yin

共享帐户跨域顺序推荐（SCSR）是一项新兴但具有挑战性的任务，它在顺序推荐中同时考虑了共享帐户和跨域特征。现有的SCSR研究主要基于递归神经网络（RNN）和图形神经网络（GNN），但它们忽略了一个事实，即虽然多个用户共享一个帐户，但一次主要由一个用户占用。这种观察激励我们通过关注其最近的行为来学习更准确的特定于用户的帐户表示。此外，尽管现有的工作为不相关的交互赋予了较低的权重，但它们仍然可能稀释域信息并阻碍跨域推荐。为了解决上述问题，我们提出了一种基于强化学习的解决方案，名为RYRL ISN，它由一个基本的跨域推荐器和一个基于强化学习的域滤波器组成。具体来说，为了对共享帐户场景中的帐户表示进行建模，基本推荐者首先将用户的混合行为聚类为潜在用户，然后利用其上的注意力模型进行用户识别。为了减少不相关域信息的影响，我们将域过滤器表述为分层强化学习任务，其中高级别任务用于决定是否修改整个传输序列，如果修改，则进一步执行低级别任务以确定是否删除其中的每个交互。为了评估我们的解决方案的性能，我们在两个真实数据集上进行了扩展实验，实验结果证明了我们的RL-ISN方法与最先进的推荐方法相比的优越性。

Shared-account Cross-domain Sequential Recommendation (SCSR) is an emergingyet challenging task that simultaneously considers the shared-account andcross-domain characteristics in the sequential recommendation. Existing workson SCSR are mainly based on Recurrent Neural Network (RNN) and Graph NeuralNetwork (GNN) but they ignore the fact that although multiple users share asingle account, it is mainly occupied by one user at a time. This observationmotivates us to learn a more accurate user-specific account representation byattentively focusing on its recent behaviors. Furthermore, though existingworks endow lower weights to irrelevant interactions, they may still dilute thedomain information and impede the cross-domain recommendation. To address theabove issues, we propose a reinforcement learning-based solution, namelyRL-ISN, which consists of a basic cross-domain recommender and a reinforcementlearning-based domain filter. Specifically, to model the account representationin the shared-account scenario, the basic recommender first clusters users'mixed behaviors as latent users, and then leverages an attention model overthem to conduct user identification. To reduce the impact of irrelevant domaininformation, we formulate the domain filter as a hierarchical reinforcementlearning task, where a high-level task is utilized to decide whether to revisethe whole transferred sequence or not, and if it does, a low-level task isfurther performed to determine whether to remove each interaction within it ornot. To evaluate the performance of our solution, we conduct extensiveexperiments on two real-world datasets, and the experimental resultsdemonstrate the superiority of our RL-ISN method compared with thestate-of-the-art recommendation methods.

4. 用于共享账户跨域顺序推荐的时间间隔增强图神经网络

Title: Time Interval-enhanced Graph Neural Network for Shared-account Cross-domain Sequential Recommendation

Published: 2022-06-16

Url: http://arxiv.org/abs/2206.08050v1

Authors: Lei Guo,Jinyu Zhang,Li Tang,Tong Chen,Lei Zhu,Hongzhi Yin

共享帐户跨域顺序推荐（SCSR）任务旨在通过利用多个域中的混合用户行为来推荐下一项。随着越来越多的用户倾向于在不同的平台上注册，并与他人共享帐户以访问特定领域的服务，它正在获得巨大的研究关注。现有的SCSR研究主要依赖于通过基于递归神经网络（RNN）的模型挖掘序列模式，这克服了以下局限性：1）基于RNN的方法主要针对发现单用户行为中的序列依赖。它们的表达能力不足以捕捉CSR中多个实体之间的关系。2）现有的方法都是通过最近空间中的知识转移来连接两个领域，而忽略了显式的跨领域图结构。3）非现有的研究考虑了项目之间的时间间隔信息，这在序列推荐中对于表征不同的项目和学习它们的判别表示至关重要。在这项工作中，我们提出了一种新的基于图的解决方案，即TiDA GCN，以解决上述挑战。具体来说，我们首先将每个域中的用户和项目链接为一个图。然后，我们使用领域感知图卷积网络来学习特定于用户的节点表示。为了充分考虑用户的特定领域偏好，我们进一步开发了两种有效的注意机制来选择性地引导消息传递过程。此外，为了进一步增强项目和帐户级别的表示学习，我们将时间间隔纳入消息传递，并设计了一个帐户感知的自注意力模块来学习项目的交互特征。实验从各个方面证明了我们提出的方法的优越性。

Shared-account Cross-domain Sequential Recommendation (SCSR) task aims torecommend the next item via leveraging the mixed user behaviors in multipledomains. It is gaining immense research attention as more and more users tendto sign up on different platforms and share accounts with others to accessdomain-specific services. Existing works on SCSR mainly rely on miningsequential patterns via Recurrent Neural Network (RNN)-based models, whichsuffer from the following limitations: 1) RNN-based methods overwhelminglytarget discovering sequential dependencies in single-user behaviors. They arenot expressive enough to capture the relationships among multiple entities inSCSR. 2) All existing methods bridge two domains via knowledge transfer in thelatent space, and ignore the explicit cross-domain graph structure. 3) Noneexisting studies consider the time interval information among items, which isessential in the sequential recommendation for characterizing different itemsand learning discriminative representations for them. In this work, we proposea new graph-based solution, namely TiDA-GCN, to address the above challenges.Specifically, we first link users and items in each domain as a graph. Then, wedevise a domain-aware graph convolution network to learn userspecific noderepresentations. To fully account for users' domainspecific preferences onitems, two effective attention mechanisms are further developed to selectivelyguide the message passing process. Moreover, to further enhance item- andaccount-level representation learning, we incorporate the time interval intothe message passing, and design an account-aware self-attention module forlearning items' interactive characteristics. Experiments demonstrate thesuperiority of our proposed method from various aspects.

5. 利用变分域不变用户嵌入进行部分重叠跨域推荐

Title: Exploiting Variational Domain-Invariant User Embedding for Partially Overlapped Cross Domain Recommendation

Published: 2022-05-27

Url: http://arxiv.org/abs/2205.06440v2

Authors: Weiming Liu,Xiaolin Zheng,Mengling Hu,Chaochao Chen

跨域推荐（CDR）已被广泛研究，以利用不同的领域知识来解决推荐系统中的冷启动问题。大多数现有的CDR模型都假设源域和目标域共享相同的重叠用户集进行知识转移。然而，在实际的CDR任务中，只有很少比例的用户同时在源域和目标域上激活。在本文中，我们重点研究了部分重叠跨域推荐（PartiallyOverlapped Cross-Domain Recommendation，POCDR）问题，即如何同时保留重叠和非重叠用户的信息以提高推荐性能。现有的方法无法充分利用跨领域非重叠用户背后的有用知识，当大多数用户都是非重叠用户时，这限制了模型的性能。为了解决这个问题，我们提出了一种具有变分域不变嵌入对齐（VDEA）模型的端到端双重自动编码器，这是一种用于POCDR问题的跨域推荐框架，它利用具有局部和全局嵌入对齐的双重变分自动编码器来利用域不变用户嵌入。VDEA首先采用变异推理来捕获协作用户偏好，然后利用Gromov-Wasserstein分布联合聚类最优传输来聚类具有相似评分交互行为的用户。我们对豆瓣和亚马逊数据集的实证研究表明，VDEA显著优于最先进的模型，尤其是在POCDR设置下。

Cross-Domain Recommendation (CDR) has been popularly studied to utilizedifferent domain knowledge to solve the cold-start problem in recommendersystems. Most of the existing CDR models assume that both the source and targetdomains share the same overlapped user set for knowledge transfer. However,only few proportion of users simultaneously activate on both the source andtarget domains in practical CDR tasks. In this paper, we focus on the PartiallyOverlapped Cross-Domain Recommendation (POCDR) problem, that is, how toleverage the information of both the overlapped and non-overlapped users toimprove recommendation performance. Existing approaches cannot fully utilizethe useful knowledge behind the non-overlapped users across domains, whichlimits the model performance when the majority of users turn out to benon-overlapped. To address this issue, we propose an end-to-enddual-autoencoder with Variational Domain-invariant Embedding Alignment (VDEA)model, a cross-domain recommendation framework for the POCDR problem, whichutilizes dual variational autoencoders with both local and global embeddingalignment for exploiting domain-invariant user embedding. VDEA first adoptsvariational inference to capture collaborative user preferences, and thenutilizes Gromov-Wasserstein distribution co-clustering optimal transport tocluster the users with similar rating interaction behaviors. Our empiricalstudies on Douban and Amazon datasets demonstrate that VDEA significantlyoutperforms the state-of-the-art models, especially under the POCDR setting.