Self-Supervised Hypergraph Convolutional Networks for Session-based Recommendation

通讯作者

主页link:https://sites.google.com/view/hongzhi-yin/home

Background

A session is a transaction with multiple purchased items in one shopping event, and SBR focuses on next-item prediction by using the real-time user.
As user profiles are generally not available in this scenario, session-based recommendation(SBR) focuses on next-item prediction at a certain time point.Hypergraph provides a natural way to capture beyond-pairwise relations, while its potential for SBR has remained unexplored.

In this paper, we fill this gap by modeling sessionbased data as a hypergraph and then propose a hypergraph convolutional network to improve SBR.

Contribution

• We propose a novel dual channel hypergraph convolutional network for SBR, which can capture the beyondpairwise relations among items through hypergraph modeling.
• We innovatively integrate a self-supervised task into the training of our network to enhance hypergraph modeling and improve the recommendation task.
• Extensive experiments show that our proposed model has overwhelming superiority over the state-of-the-art baselines and achieves statistically significant improvements on benchmark datasets.

Related Work

Hypergraph Learning

The most relevant work to ours is HyperRec (Wang et al. 2020a), which uses hypergraph to model the short-term user preference for next-item recommendation.
However, it does not exploit inter-hyperedge information and is not designed for session-based scenarios. Besides, the high complexity of this model makes it impossible to be deployed in real scenarios.

Self-supervised Learning

Self-supervised learning (Hjelm et al. 2018) is an emerging machine learning paradigm which aims to learn the data representation from the raw data.
自监督学习是无监督学习里面的一种，也被称作(pretext task)。自监督学习主要是利用辅助任务（pretext）从大规模的无监督数据中挖掘自身的监督信息，通过这种构造的监督信息对网络进行训练，从而可以学习到对下游任务有价值的表征。
Currently,the potentials of self-supervised learning for hypergraph representation learning and SBR have not been investigated. We
are the first to integrate self-supervised learning into the scenarios of SBR and hypergraph modeling.

The proposed method

首先介绍，文章中涉及到的定义:

Hypergraph Constrction

After transforming the session data into a hypergraph, any two items clicked in a session are connected.
Compared with the hypergraph which depicts the item-level high-order relations, the line graph describes the session-level relations that are also termed cross-session information

Hypergraph Convolutional Network

Hypergraph Channel and Convolution

超图卷积可以被看作是一个两阶段的细化，对超图结构进行 "结点-超边-结点 "的特征转换。转变的超图结构.
超图卷积定义如下:

行归一化，这个式子最后两项代表信息从节点传递到超边，然后再乘上的H代表信息从超边到节点。

文章还把时间信息融合进来了，于是文章合并了reversed position embeddings with the learned item representations by a learnable position matrix.于是在某个session中个每个item 的embedding 由下式表示:

一个session的embedding就是其中包含的item的embedding的平均

Model Optimization and Recommendation Generation

在这部分，文章首先为每个候选item计算一个分数，然后利用softmax函数计算该item成为下一个被推荐的物品的可能性。
文章通过最小化交叉熵作为模型优化目标。

Enhancing SBR with Self-Supervised Learning

we first initialize the channel-specific session embeddings, by looking up the items belonged to each session and then averaging the corresponding items embeddings in X
line graph convolution如下所示:

Creating self-supervision signals.

Since each channel encodes a (hyper)graph that only depicts either of the item-level (intra-session) or the session-level (inter-session) structural information of the session-induced hypergraph, the two groups of embeddings know little about each other but can mutually complement.
具体做法:
对于每个小批次，包括训练中的n个环节。在这两组sessions嵌入之间存在着一个session embedding的双射。于是这两组可以成为自监督学习的groud truth而这种一对一的映射被视为label的增强。如果两个session embeddinga都表示同一session，我们就把这一对标记为ground-truth，否则我们将其标记为negative。

Contrastive learning

参考 (Velickovic et al. 2019;Bachman, Hjelm, and Buchwalter 2019), 文章将这两个channels作为两个反映同一session的不同视图。

这个学习目标被解释为最大化在不同视图中学习的session之间的mutual information.通过这样做，它们可以相互获取信息，从而通过卷积操作来提高自己在item/session特征提取中的性能，且可以利用跨会话信息来改进它们的嵌入。

最后,我们统一推荐任务和自监督任务为下式:

Experiments

Experimental Settings

数据集：
Tmall dataset comes from IJCAI-15 competition, which contains anonymized user’s shopping logs on Tmall online shopping platform. Nowplaying dataset describes the music listening behavior of users

消融实验:
DHCN-P and DHCN-NA. DHCN-P represents the version without the reversed position embeddings, and DHCN-NA represents the
version without the soft attention mechanism.
从左图看出，没有这两部分效果更好，文章解释到可能是因为内在联系比严格的序列更重要。

Impact of Model Depth.

To study the impacts of hypergraph convolutional network’s depth in session-based recommendation, we range the numbers of layers of the network within {1, 2, 3, 4, 5}
根据图3所示的结果，DHCN对Diginetica上的层数并不十分敏感，且三层设置是最好的。

Impact of Self-Supervised Learning.

We introduce a hyper-parameter β to control the magnitude of self-supervised learning.
发现 β 可以促进 Prec@20 and MRR@20, 且随着增加性能下降。

Conclusion

1. 现有的基于GNNs的SBR模型将items转换视为成对关系，这无法捕捉到items之间无处不在的高阶关联。在本文中，提出了一种双通道超图卷积网络。
2. 为了进一步加强该网络，文章创新性地将自我监督整合到网络训练中。
3. 广泛的实证研究证明了文章的模型具有压倒性的优势。消融研究验证了超图卷积和自监督的有效性和合理性。