[论文解读] Self-Supervised Hypergraph Convolutional Networks for Session-based Recommendation
介绍了一种双通道超图卷积网络(DHCN),用于基于会话的推荐,并通过自监督学习进一步增强,采用线图通道并在两个视图之间最大化互信息,以提升推荐性能。
Session-based recommendation (SBR) focuses on next-item prediction at a certain time point. As user profiles are generally not available in this scenario, capturing the user intent lying in the item transitions plays a pivotal role. Recent graph neural networks (GNNs) based SBR methods regard the item transitions as pairwise relations, which neglect the complex high-order information among items. 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 session-based data as a hypergraph and then propose a hypergraph convolutional network to improve SBR. Moreover, to enhance hypergraph modeling, we devise another graph convolutional network which is based on the line graph of the hypergraph and then integrate self-supervised learning into the training of the networks by maximizing mutual information between the session representations learned via the two networks, serving as an auxiliary task to improve the recommendation task. Since the two types of networks both are based on hypergraph, which can be seen as two channels for hypergraph modeling, we name our model extbf{DHCN} (Dual Channel Hypergraph Convolutional Networks). Extensive experiments on three benchmark datasets demonstrate the superiority of our model over the SOTA methods, and the results validate the effectiveness of hypergraph modeling and self-supervised task. The implementation of our model is available at https://github.com/xiaxin1998/DHCN
研究动机与目标
- 识别在未可用用户画像的会话中捕捉超越成对项相关性的动机。
- 将会话数据建模为超图,以编码项之间的高阶交互。
- 提出双通道架构,包含超图和线图通道,以改进表征学习。
- 通过在两个视图之间最大化互信息来整合自监督学习,以提升推荐性能。
提出的方法
- 构建一个超图,其中每个会话形成一个超边,将该会话中的所有项连接起来。
- 使用带有节点-超边-节点细化的超图卷积来获得项嵌入。
- 引入位置嵌入,以在不进行完整序列建模的情况下注入时间信息。
- 构建超图的线图,以捕捉跨会话信息,并对会话级表示执行线图卷积。
- 引入自监督的两视图对比目标,在超图通道与线图通道表示之间最大化互信息。
- 以主要推荐损失和辅助自监督损失共同训练(L = L_r + beta * L_s)。
实验结果
研究问题
- RQ1超图建模是否比成对图更能捕捉会话中的高阶项关系?
- RQ2双通道(超图和线图)架构是否提升会话基于推荐的性能?
- RQ3在两个视图之间加入自监督对比学习是否进一步增强表示和准确性?
主要发现
| 方法 | Tmall P@10 | Tmall M@10 | Tmall P@20 | Tmall M@20 | Nowplaying P@10 | Nowplaying M@10 | Nowplaying P@20 | Nowplaying M@20 | Diginetica P@10 | Diginetica M@10 | Diginetica P@20 | Diginetica M@20 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Item-KNN | 6.65 | 3.11 | 9.15 | 3.31 | 10.96 | 4.55 | 15.94 | 4.91 | 25.07 | 10.77 | 35.75 | 11.57 |
| FPMC | 13.10 | 7.12 | 16.06 | 7.32 | 5.28 | 2.68 | 7.36 | 2.82 | 15.43 | 6.20 | 26.53 | 6.95 |
| GRU4REC | 9.47 | 5.78 | 10.93 | 5.89 | 6.74 | 4.40 | 7.92 | 4.48 | 17.93 | 7.33 | 29.45 | 8.33 |
| NARM | 19.17 | 10.42 | 23.30 | 10.70 | 13.60 | 6.62 | 18.59 | 6.93 | 35.44 | 15.13 | 49.70 | 16.17 |
| STAMP | 22.63 | 13.12 | 26.47 | 13.36 | 13.22 | 6.57 | 17.66 | 6.88 | 33.98 | 14.26 | 45.64 | 14.32 |
| SR-GNN | 23.41 | 13.45 | 27.57 | 13.72 | 14.17 | 7.15 | 18.87 | 7.47 | 36.86 | 15.52 | 50.73 | 17.59 |
| FGNN | 20.67 | 10.07 | 25.24 | 10.39 | 13.89 | 6.80 | 18.78 | 7.15 | 37.72 | 15.95 | 50.58 | 16.84 |
| DHCN | 25.14* | 13.91* | 30.43* | 14.26* | 17.22 | 7.78 | 23.03 | 8.18 | 39.87 | 17.53 | 53.18 | 18.44 |
| S2-DHCN | 26.22 | 14.60 | 31.42 | 15.05 | 17.35 | 7.87 | 23.50 | 8.18 | 40.21 | 17.59 | 53.66 | 18.51 |
- DHCN在三个基准数据集上的P@10, M@10, P@20, MRR@20指标上超过最先进基线。
- 自监督版本(S2-DHCN)在DHCN基础上带来额外提升,表明辅助任务有效。
- 超图建模提供了超越对项关系的表示;线图通道捕捉跨会话信息,补充项级建模。
- 消融研究显示位置嵌入和软注意力的贡献,且对数据集有不同影响。
- 自监督学习对短会话尤其有益,因为数据稀疏性更高。
更好的研究,从现在开始
从论文设计到论文写作,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。