Enhancing session-based social recommendation through item graph embedding and contextual friendship modeling
Enhancing session-based social recommendation through item graph embedding and contextual friendship modeling
复制标题
通过项目图嵌入和上下文友谊建模增强基于会话的社交推荐
DOI：
10.1016/j.neucom.2020.08.023
复制
发表时间：
2021-01-02
期刊：
影响因子：
6
通讯作者：
Wu, Jian
中科院分区：
文献类型：
--
作者：
Gu, Pan;Han, Yuqiang;Wu, Jian
推荐系统旨在帮助用户从在线平台上的大量候选人中找到匹配的项目。在许多在线平台中，例如Yelp和Epinions，用户的行为会随着时间的推移被不断记录，用户还可以与他人建立联系并分享他们的兴趣。以前的推荐方法要么对动态兴趣进行建模，要么对动态社会影响进行建模。一些研究集中在这两个因素的建模上，但它们仍然有一些局限性：1）他们没有考虑所有会话序列之间复杂的项目转换，这可以用作提高推荐方法性能的局部因素，2）他们通过在时间 t 保持所有目标用户的朋友向量不变，忽略了用户和他们的朋友仅在某些会话中共享相同的偏好，3）他们没有考虑用户的长期偏好可能会随着时间的演变而改变。为了克服上述问题，在本文中，我们提出了一种将项目图嵌入和上下文友谊建模结合到推荐任务中的方法。具体来说，1）我们基于所有历史会话序列构建有向项目图，并利用图神经网络来捕获项目之间丰富的局部依赖关系，2）采用会话级注意力机制根据目标用户当前的兴趣获取每个朋友的表示，3）对目标用户的历史会话兴趣应用最大池化以了解他/她的长期兴趣的动态。对两个现实世界数据集的广泛实验表明，我们提出的模型在各种评估指标上始终优于最先进的方法。 (c) 2020 Elsevier B.V. 保留所有权利。
Recommender systems are designed to help users find matching items from plenty of candidates in online platforms. In many online platforms, such as Yelp and Epinions, users' behaviors are constantly recorded over time, and the users also can build connections with others and share their interests. Previous recommendation methods have either modeled the dynamic interests or the dynamic social influences. A few studies have focused on the modeling of both factors, but they still have several limi-tations: 1) they fail to consider the complex items transitions among all session sequences, which can be used as a local factor to boost the performance of recommendation methods, and 2) they ignore that a user and their friends only share the same preferences in certain sessions, by keeping the friend vector unchanged for all target users at time t, and 3) they do not consider that a user's long-term preference may change with the evolution of interests.To overcome the above issues, in this paper, we propose an approach to incorporate item graph embedding and contextual friendship modeling into the recommendation task. Specifically, 1) we construct a directed item graph based on all historical session sequences and utilize a graph neural network to capture the rich local dependency between items, and 2) take a session-level attention mechanism to get each friend's representation according to the target user's current interests, and 3) apply max-pooling on the target user's historical session interests to learn the dynamics of his/her long-term interests. Extensive experiments on two real-world datasets show that our proposed model outperforms stateof-the-art methods consistently on various evaluation metrics. (c) 2020 Elsevier B.V. All rights reserved.