[Paper Review] Self-Supervised Hypergraph Convolutional Networks for Session-based Recommendation

Xin Xia, Hongzhi Yin+4 more|arXiv (Cornell University)|Dec 12, 2020

Recommender Systems and TechniquesComputer Science43 references34 citations

TL;DR

Introduces a dual-channel hypergraph convolutional network (DHCN) for session-based recommendation and enhances it with self-supervised learning via a line-graph channel and mutual information maximization between the two views.

ABSTRACT

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

Motivation & Objective

• Motivate capturing beyond-pairwise item correlations in sessions where user profiles are unavailable.
• Model session data as a hypergraph to encode high-order interactions among items.
• Propose a dual-channel architecture with hypergraph and line-graph channels to improve representation learning.
• Integrate self-supervised learning by maximizing mutual information between the two views to boost recommendation performance.

Proposed method

• Construct a hypergraph where each session forms a hyperedge linking all items in that session.
• Apply hypergraph convolution with a node-hyperedge-node refinement to obtain item embeddings.
• Incorporate position embeddings to inject temporal information without full sequence modeling.
• Build a line graph of the hypergraph to capture cross-session information and perform line-graph convolution on session-level representations.
• Introduce a self-supervised two-view contrastive objective that maximizes mutual information between the hypergraph-channel and the line-graph-channel representations.
• Train jointly with a primary recommendation loss and an auxiliary self-supervised loss (L = L_r + beta * L_s).

Experimental results

Research questions

• RQ1 Can hypergraph modeling capture high-order item relations in sessions better than pairwise graphs?
• RQ2 Does a dual-channel (hypergraph and line graph) architecture improve session-based recommendation performance?
• RQ3 Does incorporating self-supervised contrastive learning between the two views further enhance representations and accuracy?

Key findings

• DHCN outperforms state-of-the-art baselines on three benchmark datasets across P@10, M@10, P@20, and MRR@20 metrics.
• Self-supervised version (S2-DHCN) yields additional gains over DHCN, indicating the effectiveness of the auxiliary task.
• Hypergraph modeling provides beyond-pairwise item relations; the line graph channel captures cross-session information to complement item-level modeling.
• Ablation studies show contributions of position embeddings and soft attention, with dataset-dependent effects.
• Self-supervised learning is particularly beneficial for shorter sessions where data sparsity is higher.