PaperlayerView exact revision on arXiv

Source-linked AI summary

Double-Scale Self-Supervised Hypergraph Learning for Group Recommendation

Junwei Zhang, Min Gao, Junliang Yu, Lei Guo, Jundong Li, Hongzhi Yin

arXiv:2109.04200v2cs.IRcs.AI

TL;DR

Group recommendation must model complex correlations among members despite pairwise interaction assumptions and sparse group-item data. The paper proposes hierarchical hypergraph learning with double-scale node dropout self-supervision, and reports superiority over state-of-the-art models on three public datasets.

  • Problem

    Existing group recommenders mainly use heuristic or attention-based aggregation, miss high-order intra- and inter-group interactions, and face sparse group-item interaction data.

  • Method

    The framework uses user- and group-level hierarchical hypergraph convolution to model tuplewise user correlations and double-scale node dropout to generate self-supervision signals.

  • Results

    Experiments on three public datasets report that HHGR and S2-HHGR outperform other state-of-the-art group recommendation models.

  • Takeaways & Limitations

    Hypergraph modeling captures user interactions within and beyond groups, while double-scale self-supervision enhances representations against data sparsity.

  • Takeaways & Limitations

    The task formulation assumes item representations are randomly initialized as d-dimensional vectors.

Abstract

from arXiv · show

With the prevalence of social media, there has recently been a proliferation of recommenders that shift their focus from individual modeling to group recommendation. Since the group preference is a mixture of various predilections from group members, the fundamental challenge of group recommendation is to model the correlations among members. Existing methods mostly adopt heuristic or attention-based preference aggregation strategies to synthesize group preferences. However, these models mainly focus on the pairwise connections of users and ignore the complex high-order interactions within and beyond groups. Besides, group recommendation suffers seriously from the problem of data sparsity due to severely sparse group-item interactions. In this paper, we propose a self-supervised hypergraph learning framework for group recommendation to achieve two goals: (1) capturing the intra- and inter-group interactions among users; (2) alleviating the data sparsity issue with the raw data itself. Technically, for (1), a hierarchical hypergraph convolutional network based on the user- and group-level hypergraphs is developed to model the complex tuplewise correlations among users within and beyond groups. For (2), we design a double-scale node dropout strategy to create self-supervision signals that can regularize user representations with different granularities against the sparsity issue. The experimental analysis on multiple benchmark datasets demonstrates the superiority of the proposed model and also elucidates the rationality of the hypergraph modeling and the double-scale self-supervision.

1 INTRODUCTION

Group recommendation must reconcile changing member influences while modeling complex user and group interactions. The paper addresses pairwise-modeling and sparsity limits with hierarchical hypergraphs and double-scale self-supervision.

  • Motivation: Group recommendation seeks item suggestions that reach consensus among members whose influences can vary across groups.The introduction illustrates how different group compositions and member positions can change the recommended items.
  • Limitations of Existing Methods: Heuristic aggregation rules ignore user interactions, while attention-based models focus on pairwise connections and miss high-order interactions within and beyond groups.The cited approaches include average, least misery, maximum satisfaction, and attention over user-item graphs.
  • Data Sparsity: Sparse group-item interaction records make it difficult to learn accurate group preferences directly from interaction data.The paper identifies sparsity as especially problematic because many groups are formed temporarily.
  • Proposed Approach: The proposed framework uses a hierarchical hypergraph to capture intra- and inter-group user interactions and double-scale self-supervision to alleviate sparsity.The framework combines user- and group-level modeling with self-supervision at different granularities.
  • Evaluation: Experiments on three group recommendation datasets report superiority over recent baselines and examine the contributions of hypergraph learning and double-scale self-supervision.The paper states that the code implementation is publicly released.

2 PRELIMINARIES

The preliminaries define hypergraphs and the group recommendation task using user, item, and group interaction structures. They also introduce the mathematical representations used for users, items, and groups.

  • 2.1 Definition of Hypergraph: A hypergraph G = (V, E) contains vertices and hyperedges, with each hyperedge connecting multiple vertices.Its incidence matrix H records whether each vertex belongs to each hyperedge.
  • 2.1 Definition of Hypergraph: The hypergraph incidence matrix H ∈ R^M×N records vertex–hyperedge membership using binary entries.The supplied definition states h_vε = 1 when vertex v belongs to hyperedge ε and 0 otherwise.
  • 2.2 Task of Group Recommendation: Group recommendation predicts an item ranking list for a group containing multiple members.The task includes user-item, group-item, and user-group interaction behaviors.
  • 2.2 Task of Group Recommendation: User-item and group-item interactions are represented by matrices R and S, whose binary entries indicate consumption.The formulation also defines user, item, and group representations in a d-dimensional embedding space.

3 THE PROPOSED METHOD

The proposed framework combines hierarchical hypergraph convolution with double-scale self-supervision to model high-order user interactions and address sparse group-recommendation data.

  • Hierarchical hypergraph: HHGR propagates information through user- and group-level hypergraphs to capture tuplewise user interactions within and beyond groups.The user-level hypergraph models member relations, while the group-level hypergraph models informative inter-group relations.
  • Hierarchical hypergraph: User-level hypergraph convolution aggregates information through incidence-defined user–hyperedge relations and normalized degree matrices.The convolution learns hyperedge representations from node features and integrates related hyperedge information into each user representation.
  • Hierarchical hypergraph: Group representations are generated by weighted aggregation of member representations before group-level hypergraph propagation.The weights reflect users’ contributions to group decision-making, and the resulting group embeddings are used to estimate item preferences.
  • Hierarchical hypergraph: The group-level hypergraph uses triadic motifs among projected group nodes to select informative inter-group interactions.Groups sharing users are connected in the projected graph, and qualifying triples form motif-induced hyperedges.
  • Double-scale self-supervision: Double-scale node dropout augments sparse raw data with coarse- and fine-grained self-supervision signals for user and group representations.The coarse strategy drops users across their associated group hyperedges, while the final user representation combines the two granularities.
  • Model optimization: S2-HHGR jointly optimizes recommendation and self-supervised objectives, first pretraining user representations and then updating supervised user–item and group–item losses.The overall objective unifies self-supervised learning with group recommendation, with the learned user embeddings contributing to group representations.

4 EXPERIENCE AND RESULTS

The experiments are designed to evaluate comparative performance, component benefits, and hyperparameter sensitivity of the proposed group-recommendation model.

  • Research questions: The experiments compare the proposed model with state-of-the-art group recommendation models.This evaluation addresses RQ1, which asks how the model performs relative to existing methods.
  • Research questions: Ablation-style analysis examines the benefits of the hierarchical hypergraph and self-supervised learning components.This evaluation addresses RQ2 by isolating the contributions of the two principal components.
  • Research questions: The experiments assess how hyperparameters influence S2-HHGR effectiveness.This evaluation addresses RQ3 through hyperparameter sensitivity analysis.

4.1 Experimental Settings

Experiments use three public group-recommendation datasets, established baselines, ranking metrics, and fixed training settings with a 70%/10%/20% group split.

  • Datasets and splits: The evaluation uses Weeplaces, CAMRa2011, and Douban, with groups randomly split into 70% training, 10% validation, and 20% test sets.The datasets represent check-in, household movie-rating, and activity-based group data, respectively.
  • Baselines: The comparison includes popularity, NeuMF, AGREE, MoSAN, SIGR, HHGR, and S2-HHGR.HHGR is the vanilla model, while S2-HHGR is its self-supervised version.
  • Metrics: Performance is measured with NDCG@K and Recall@K for K = {20, 50}.NDCG@K evaluates ranking of true items, while Recall@K measures retrieved relevant items among the top K.
  • Implementation settings: General settings use embedding size 64, batch size 512, and 10 negative samples.The self-supervised learning rate is 5e-4, and the group-level hypergraph learning rate is 1e-4.

4.2 Recommendation Performance Comparison (RQ1)

The experiments compare HHGR and S2-HHGR with baseline group recommenders across three datasets and under sparse interaction conditions. S2-HHGR achieves the strongest reported performance, supporting the value of hypergraph modeling and double-scale self-supervision.

  • Attention-based group models outperform Popular and NeuMF on most datasets by modeling user interactions and assigning different member weights.
  • HHGR outperforms most attention-based models, including AGREE, MoSAN, and SIGR, indicating benefits from high-order interaction modeling.
  • S2-HHGR beats the most advanced group recommendation models on all three datasets, although HHGR is slightly inferior to GroupIM.
  • NDCG and Recall generally increase as an item receives interactions from more groups, because additional interactions provide more information for learning group preferences.
  • On the Weeplaces sparsity analysis, S2-HHGR shows the best performance compared with the other evaluated models.

4.3 Ablation Study (RQ2)

Ablation studies examine the hierarchical hypergraph and the two self-supervised node-dropping granularities. Removing either hypergraph level or self-supervision reduces performance, while moderate dropping strengths work best.

  • Investigation of the hierarchical hypergraph: Removing either the user-level or group-level hypergraph reduces performance on Weeplaces and CAMRa2011 relative to the complete HHGR model.HHGR-wg removes the group level, whereas HHGR-wu removes the user level.
  • Investigation of self-supervised learning: S2-HHGR performs best among the independent-granularity variants, while HHGR-F generally exceeds HHGR-C.The authors attribute the stronger fine-grained contribution partly to its generation of more hard samples.
  • Investigation of self-supervised learning: Without self-supervision, HHGR declines in most cases on the two evaluated datasets, indicating that the self-supervised strategy contributes to recommendation performance.
  • Investigation of self-supervised learning: Fine-grained and coarse-grained node-dropping rates peak at approximately 0.3 and 0.2, respectively, before performance decreases.

4.4 Parameter Sensitivity Analysis (RQ3)

Parameter sensitivity is evaluated on Weeplaces across learning rate, convolutional depth, batch size, and negative-sample count. Performance peaks at selected intermediate settings and declines with excessive convolutional depth.

  • A hypergraph convolutional depth of 2 produces the best performance, while deeper networks steadily decline.The authors suggest that multilayer hypergraph convolution can cause over-smoothing.
  • The optimal batch size is 512 in the reported Weeplaces sensitivity analysis.
  • The model achieves its best Weeplaces performance with 10 negative samples.

5 RELATED WORK

Early group recommenders aggregate member preferences heuristically, while later attention-based models learn pairwise user interactions. Hypergraph and self-supervised approaches address higher-order structure and sparse recommendation data, with this work combining both through double-scale dropout.

  • Preference aggregation: Early methods aggregate group preferences using average, least misery, or maximum satisfaction strategies.These strategies respectively target mean preference, the least satisfied member, or maximum satisfaction.
  • Attention-based models: Attention-based models learn users’ differing influences and intra-group interactions, but remain focused on pairwise connections.Examples include GAME and other attention-based approaches over heterogeneous or user-item graphs.
  • Hypergraph learning: Graph neural networks capture high-order relations but are limited by pairwise connectivity, motivating hypergraphs for complex higher-order structures.Hypergraph recommender studies apply this representation to model complex relations beyond ordinary graphs.
  • Hypergraph learning: This work differs from prior hypergraph representation learning by using motifs to select informative group interactions.The stated distinction is motif-based interaction selection rather than only adopting hypergraph representation learning.
  • Self-supervised recommendation: Self-supervised recommendation methods generate additional signals through generative reconstruction or contrastive positive-negative representation learning.Prior examples include pretraining strategies in S3-Rec and masked-item prediction in Bert4Rec.
  • Self-supervised recommendation: Unlike GroupIM, this work uses double-scale node dropout across different hypergraph views to enhance representations and recommendation performance.GroupIM maximizes mutual information between users and groups, whereas this approach maximizes mutual information between views of the same hypergraph.

6 CONCLUSION AND FUTURE WORK

The paper proposes HHGR and its self-supervised extension S2-HHGR, combining hierarchical hypergraph convolution with double-scale node dropout for group recommendation. Experiments on three public datasets report superiority over state-of-the-art models, while future work targets broader self-supervised auxiliary tasks.

  • Conclusion: HHGR uses hierarchical hypergraph convolution and self-supervised learning to capture inter- and intra-group user interactions while alleviating data sparsity.The framework addresses both interaction modeling and sparse group-recommendation data.
  • Conclusion: Triadic motifs provide user interactions beyond groups for more reliable user interactions and more accurate group representations.The conclusion identifies motif-based interaction modeling as a source of improved group representation quality.
  • Experimental evaluation: Experiments on three public datasets verified HHGR and S2-HHGR’s superiority over other state-of-the-art models.S2-HHGR is the enhanced version incorporating self-supervised learning.
  • Future work: Future work will deepen self-supervised learning and design more general recommendation auxiliary tasks.The stated goal is to improve recommendation performance through broader auxiliary-task design.
Loading 2109.04200v2…