学术讲座

当前位置: 首页 - 学术讲座 - 正文

人工智能学院2021年系列学术活动——澳大利亚麦考瑞大学博士王琪学术报告

发布时间:2021-06-21点击:


人工智能学院2021年系列学术活动——澳大利亚麦考瑞大学博士王琪学术报告

 

 

 

报告题目:基于用户和物品三元关系的社交推荐

Title: Capturing the Tri-relationships among Users and Items for Social Recommendation.

报告内容:社交推荐利用社交信息来缓解传统基于协同过滤的推荐方法所面临的数据稀疏和冷启动问题。但是, 现有的社交推荐方法主要依赖于已建立的用户的社交关系和用户的评分,而没有考虑隐藏在社交媒体数据中的用户和物品之间的丰富关系。总的来说, 现有的工作有以下几个局限性:(1)在用户特征建模中,它们无法捕捉到来自与信任/朋友链接不直接相关的用户的隐含影响;同样,(2)在项目特征建模中没有很好地考虑复杂的项目-项目隐式关系来挖掘潜在的影响因素; (3) 大多数现有工作都忽略了丰富的用户-项目交互行为,例如可以反映用户偏好并揭示物品特征的评论信息 最近,图神经网络 (GNN) 已被证明可有效提高推荐性能,因为它们能够为复杂的图数据学习提供更有意义的表示。特别是图神经网络的扩展,图注意力网络 (GAT),在社交推荐方面取得了巨大成功。因此本报告介绍一种新的基于多图注意力网络的社交推荐方法来捕获用户和物品之间复杂的三元关系以实现更好的社交推荐其中, 设计的多图注意力网络是学习用户-用户、用户-项目和项目-项目之间复杂关系的核心两个真实数据集的大量实验证明了所提出模型的有效性。 

Abstract: Social recommendation leverages social information to alleviate data-sparsity and cold-start problems suffered in traditional collaborative filtering based recommendation methods. Existing social recommendation methods mainly rely on the established user's social relations and user's ratings without considering rich relations among users and items that are hidden in the social media data. These methods have several limitations: (1) they fail to capture the implicit influence from the users that are not directly connected with trust/friend links in user feature modeling; Similarly, (2) the complicated implicit item-item relations have not been well-considered to mine the influencing latent factors in item feature modeling; (3) most existing works ignore the rich user-item interactions such as reviews that can reflect user's preference and reveal item's features as well. Recently, Graph Neural Networks (GNNs) have been proved to be effective for advancing recommendation performance as they are capable of learning meaningful representations for complicated graph data. Particularly, an extension of GNNs, Graph Attention Networks (GATs), have shown great success in social recommendation. In this work, we propose a novel multi-Graph Attention Network based approach for social recommendation. The designed multi-graph attention network is the working horse to learn the complex and complicated relations between/among user-user, user-item, and item-item. Extensive experiments against two real-world datasets demonstrate the effectiveness of the proposed model.

报告人介绍:王琪,女,麦考瑞大学计算机博士,本科和研究生毕业于吉林大学计算机科学与技术学院。主要研究方向:在线社交网络中的信任预测、 社会数据分析与挖掘 、图挖掘和社会推荐等。

报告时间:2021年622 星期二 上午910-1010

报告平台:腾讯会议

会议号:803 165 285

会议密码:202106