PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-Supervised Hypergraph Convolutional Networks for Session-based Recommendation

Xin Xia, Hongzhi Yin, Junliang Yu, Qinyong Wang, Lizhen Cui, Xiangliang Zhang

arXiv:2012.06852v5cs.IR

TL;DR

Session-based recommendation must predict next items without long-term user profiles, while pairwise-transition models overlook high-order item correlations. DHCN models sessions with dual hypergraph-based channels and mutual-information self-supervision, and experiments report superiority over state-of-the-art methods.

Abstract

from arXiv · show

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 \textbf{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

Introduction

Session-based recommendation must infer intent without long-term user profiles, but sequence- and pairwise-transition models miss item coherence and high-order correlations. DHCN addresses this with dual-channel hypergraph modeling and self-supervised learning, achieving statistically significant benchmark improvements.

  • SBR predicts the next item from real-time session behavior, usually without available long-term user profiles.Source: Introduction · p. 1
  • RNN sequence models can impose overly strict order assumptions, while GNNs represent transitions pairwise and neglect complex high-order item correlations.Source: Introduction · p. 1
  • DHCN models each session as a hyperedge, capturing beyond-pairwise relations while emphasizing item coherence rather than strict relative order.Source: Introduction · pp. 1–2
  • The model uses hypergraph and line-graph channels to represent item-level and session-level information, respectively.Source: Introduction · p. 2
  • Self-supervised learning maximizes mutual information between the two channel representations as an auxiliary task for recommendation.Source: Introduction · p. 2
  • Extensive experiments report statistically significant improvements over state-of-the-art baselines on benchmark datasets.Source: Introduction · p. 2

Related Work

Prior SBR research progressed from Markov and recurrent sequence models to graph neural networks, but existing approaches still miss higher-order item correlations. Hypergraph and self-supervised recommendation research provide relevant foundations, while their combination for SBR remains underexplored.

  • Early SBR work used Markov decision processes, followed by deep recurrent and convolutional models for sequential data.Source: Session-based Recommendation · p. 2
  • GNN-based SBR methods learn item transitions over session-induced graphs, yet existing studies fail to capture complex higher-order item correlations.Source: Session-based Recommendation · p. 2
  • Hypergraph neural networks apply convolution to hypergraphs, and later work introduced dynamic and line-hypergraph convolutional models.Source: Hypergraph Learning · p. 2
  • HyperRec models short-term preference with hypergraphs but omits inter-hyperedge information and is not designed for session-based scenarios.Source: Hypergraph Learning · p. 2
  • Self-supervised learning learns representations from raw data, with contrastive graph methods using mutual-information maximization between congruent and incongruent views.Source: Self-supervised Learning · pp. 2–3
  • The potential of self-supervised learning for hypergraph representation learning and SBR had not been investigated in the cited related work.Source: Self-supervised Learning · p. 3

The Proposed Method

The proposed method models session data as a hypergraph, applies hypergraph convolution for SBR, and adds a line-graph channel with self-supervised learning to enhance hypergraph modeling.

  • The method introduces session-to-hypergraph modeling, hypergraph convolution, a line-graph channel, and integrated self-supervised learning.Source: The Proposed Method · p. 3

Notations and Definitions

The paper defines item and session representations, hypergraphs through vertices, weighted hyperedges, and incidence matrices, and line graphs through shared hyperedge nodes.

Hypergraph Construction

DHCN represents each session as a hyperedge in a hypergraph, capturing item-level high-order relations rather than only pairwise transitions. The resulting line graph captures cross-session relations between sessions.

Hypergraph Convolutional Network

The hypergraph channel propagates item information through session hyperedges and combines refined item representations with reversed positional information. Soft attention then aggregates session items for next-item prediction.

Convolution

  • Hypergraph convolution performs node-hyperedge-node feature transformation to propagate information through the hypergraph.Information first aggregates from nodes to hyperedges, then from hyperedges back to nodes.Source: Hypergraph Channel and Convolution · p. 4
  • The model averages item embeddings from the initial representation and all L convolutional layers to obtain final item embeddings.Source: Hypergraph Channel and Convolution · p. 4

Position information

  • Reversed position embeddings are integrated with learned item representations, providing the model’s only temporal factor.The model omits sequence mechanisms such as GRU units and self-attention.Source: Hypergraph Channel and Convolution · p. 4

Session representation

Recommendation

  • The hypergraph channel scores candidate items by inner product with the session embedding, then applies softmax to produce next-item probabilities.Training minimizes a cross-entropy recommendation loss.Source: Model Optimization and Recommendation… · p. 4

Enhancing SBR with Self-Supervised Learning

DHCN adds a line-graph channel to provide cross-session information and uses contrastive self-supervision to align the two channel views. The recommendation and auxiliary objectives are optimized jointly.

Line graph channel

Dual views

Self-supervision

  • Matching embeddings for the same session across channels are positives, while mismatched pairs are labeled negatives for self-supervision.The one-to-one correspondence between channel outputs supplies the labels.Source: Creating self-supervision signals · p. 5

Contrastive learning

  • Contrastive learning uses a noise-contrastive BCE objective to maximize agreement between session representations from the two views.Negative samples are created by row-wise and column-wise shuffling, and agreement is scored by a discriminator.Source: Contrastive learning · p. 5

Effect

  • The auxiliary self-supervised task can refine embeddings for sessions containing only a few items by leveraging cross-session information.Source: Contrastive learning · p. 5

Joint learning

  • The joint objective combines recommendation loss and self-supervised loss as L = Lr + βLs.β controls the magnitude of the self-supervised task.Source: Contrastive learning · p. 5

Experimental Settings

The study evaluates DHCN on Tmall, Nowplaying, and Diginetica using standard session filtering, train/test splits, representative baselines, and P@K and MRR@K metrics. It reports overall comparisons and uses dataset-specific layer settings.

Datasets

  • Experiments use the real-world benchmark datasets Tmall, Nowplaying, and Diginetica.Tmall contains anonymized shopping logs, while Nowplaying records music-listening behavior.Source: Datasets · p. 5

Preprocessing

  • Sessions with one item and items appearing fewer than five times are removed from Tmall and Nowplaying.Source: Datasets · p. 5

Baselines

Results

  • Table 2 reports overall performance for DHCN variants and comparison methods across the three datasets.S2-DHCN denotes the self-supervised version, and improvements are computed against the best baseline.Source: Table 2 · p. 6

Metrics

  • Recommendation quality is evaluated with P@K and MRR@K.These denote Precision and Mean Reciprocal Rank at K.Source: Evaluation Metrics · p. 6

Configuration

  • DHCN uses three convolutional layers for Nowplaying and Diginetica, but one layer for Tmall.The general embedding size is 100, with batch size 100 and L2 regularization 10^-5.Source: Hyper-parameters Settings · p. 6

Experimental Results

DHCN outperforms baseline methods across datasets, while ablations show benefits from hypergraph modeling and self-supervised learning. Performance also depends on dataset characteristics, network depth, and self-supervision strength.

Overall Performance

  • GNN-based SR-GNN and FGNN outperform RNN-based models, but their gains are trivial compared with DHCN.The comparison attributes the improvement of GNN-based models to their capacity for graph representation.Source: Overall Performance · p. 6
  • DHCN shows overwhelming superiority over all baselines on all datasets.The model captures beyond-pairwise relations through hypergraphs and uses limited parameters in its two convolutional channels.Source: Overall Performance · p. 6

Component Contributions

  • Self-supervised learning provides decent additional improvements, especially on the two datasets with shorter average session lengths.The results are consistent with the assumption that session sparsity can limit hypergraph modeling benefits.Source: Overall Performance · p. 6
  • On Tmall, removing reversed position embeddings or soft attention improves both metrics over the full model, whereas Diginetica shows a different pattern.The Tmall result is presented as supporting the importance of coherence over strict order modeling in that setting.Source: Ablation Study · p. 7

Impact of Model Depth

  • DHCN is best with three layers on Diginetica and one layer on Tmall; increasing depth reduces Tmall MRR@20.The reported possible cause is increasingly over-smoothed item representations.Source: Impact of Model Depth · p. 7

Impact of Self-Supervised Learning

  • Smaller self-supervision weights β improve Prec@20 and MRR@20 on both datasets, while larger β values reduce performance.The paper attributes the decline to gradient conflicts between the recommendation and self-supervised tasks, with larger β especially harming MRR@20.Source: Impact of Self-Supervised Learning · p. 7

Conclusion

The paper addresses the inability of pairwise GNN-based SBR models to capture high-order item correlations by proposing a dual-channel hypergraph convolutional network with self-supervised learning. Experiments report strong performance and validate both hypergraph convolution and the auxiliary task.

Conclusion

  • DHCN models session data with hypergraph convolution to capture high-order item correlations beyond pairwise relations.The model additionally integrates self-supervised learning to enhance its network.Source: Conclusion · p. 7
  • Extensive empirical studies demonstrate DHCN’s overwhelming superiority, while ablations validate hypergraph convolution and self-supervised learning.Source: Conclusion · p. 7

Related papers

Loading 2012.06852v5…