Source-linked AI summary
Hierarchical Hyperedge Embedding-based Representation Learning for Group Recommendation
Lei Guo, Hongzhi Yin, Tong Chen, Xiangliang Zhang, Kai Zheng
TL;DR
Occasional group recommendation must infer group preferences when group-item histories are limited, even though member preferences may not represent the group and user-item data are sparse. HyperGroup combines social preference modeling, overlapping-group hyperedge embeddings, and joint training; it significantly outperforms the baselines on two datasets. Its evaluation is limited to datasets suitable for the method’s social-network setting.
Problem
Occasional group recommendation has limited group-item histories, while member preferences may differ from group preferences and user-item interactions are also sparse.
Source: §1 · p. 3Method
HyperGroup uses a GNN to learn individual preferences from social neighbors, embeds groups as hyperedges linked by shared members, and jointly trains user-item and group-item tasks.
Source: §1 · p. 3Results
HyperGroup significantly outperforms all baselines on both datasets, with all improvements statistically significant at p<0.01.
Source: §5 · p. 17Takeaways & Limitations
The experiments on two real-world datasets demonstrate HyperGroup’s superiority over the baselines for occasional group recommendation.
Source: §1 · p. 4The evaluation excludes CAMRa2011 and Movielens-Group because they lack social-network information or contain persistent or randomly generated groups.
Source: §4.2 · p. 15
Abstract
from arXiv · showhide
In this work, we study group recommendation in a particular scenario, namely Occasional Group Recommendation (OGR). Most existing works have addressed OGR by aggregating group members' personal preferences to learn the group representation. However, the representation learning for a group is most complex beyond the fusion of group member representation, as the personal preferences and group preferences may be in different spaces. In addition, the learned user representation is not accurate due to the sparsity of users' interaction data. Moreover, the group similarity in terms of common group members has been overlooked, which however has the great potential to improve the group representation learning. In this work, we focus on addressing the above challenges in group representation learning task, and devise a hierarchical hyperedge embedding-based group recommender, namely HyperGroup. Specifically, we propose to leverage the user-user interactions to alleviate the sparsity issue of user-item interactions, and design a GNN-based representation learning network to enhance the learning of individuals' preferences from their friends' preferences, which provides a solid foundation for learning groups' preferences. To exploit the group similarity to learn a more accurate group representation from highly limited group-item interactions, we connect all groups as a network of overlapping sets, and treat the task of group preference learning as embedding hyperedges in a hypergraph, where an inductive hyperedge embedding method is proposed. To further enhance the group-level preference modeling, we develop a joint training strategy to learn both user-item and group-item interactions in the same process. We conduct extensive experiments on two real-world datasets and the experimental results demonstrate the superiority of our proposed HyperGroup in comparison to the state-of-the-art baselines.
1 INTRODUCTION
Occasional group recommendation is difficult because group histories are sparse, member preferences may not capture group preferences, and user interactions are also sparse. HyperGroup combines social preference modeling, group-overlap hyperedge learning, and joint training to address these challenges.
1 INTRODUCTION
- Occasional groups may have little or no historical group-item feedback, and members’ differing social influence makes their decisions harder to model than individual preferences.Persistent groups instead have stable members and abundant group-item interactions.Source: §1 · p. 2
- Group preferences may differ from members’ personal preferences, while common members can signal similarity between groups and help address sparse group-item data.Existing methods also rely on individual activity data, although user-item interactions are themselves often extremely sparse.Source: §1 · pp. 2–3
- HyperGroup models groups as hyperedges in a network of overlapping groups, learning each group representation from representations of groups sharing at least one member.Its components are Individual Preference Modeling, Hyperedge Representation Learning, and joint training.Source: §1 · p. 3
- Its Individual Preference Modeling component uses a GNN over user-user interactions to learn members’ preferences from friends’ preferences, addressing sparse user-item interactions.These learned individual preferences provide a foundation for group preference learning.Source: §1 · p. 3
- HyperGroup jointly trains user-item and group-item recommendation tasks, and experiments on two real-world datasets report superiority over baselines.The authors distinguish it from methods that model each group individually by exploiting group similarity through common members.Source: §1 · pp. 3–4
2 RELATED WORK
Group recommendation research distinguishes persistent groups with plentiful feedback from occasional groups with limited or unavailable feedback. Prior occasional-group approaches aggregate member preferences or predictions, while this work highlights social homophily and overlap-based group similarity as additional modeling signals.
2.2 Occasional Group Recommendation
- Occasional-group recommendation must learn group preferences from member preferences because group feedback is highly limited or unavailable.The related-work section divides these methods into late and early aggregation approaches.Source: §2 · p. 4, §2.2 · p. 4
2.1 Persistent Group Recommendation
- Persistent-group methods often treat groups as pseudo-users because stable membership and rich interaction histories support conventional personalized recommendation techniques.Examples include methods requiring known group preferences or substantial group behavior history, limiting their use to predefined groups.Source: §2.1 · p. 4
2.2.1 Late Aggregation methods.
- Late aggregation first predicts for individual members, then combines their scores or lists using strategies such as average satisfaction, least misery, or maximum pleasure.These predefined heuristics are described as unstable and sub-optimal, with none best across all datasets.Source: §2.2.1 · pp. 4–5
2.2.2 Early Aggregation methods.
- Early aggregation combines members’ preferences or representations into a group profile before making recommendations.Prior work includes probabilistic models, attention-based aggregation, social influence, and group self-attention approaches.Source: §2.2.2 · p. 5
- The paper distinguishes HyperGroup by modeling user preferences from neighbors and group preferences through hyperedge embeddings weighted by overlapping groups and their connecting users.The related-work discussion notes that prior methods did not jointly study group-group correlation and social homophily in the user network.Source: §2.2.2 · p. 5
3 METHODOLOGIES
HyperGroup addresses sparse occasional group recommendation by learning socially enhanced user preferences and group embeddings from overlapping groups, using user, item, group, social, interaction, and group-overlap information. It jointly optimizes user-item and group-item interactions with stochastic gradient descent and negative sampling.
3.1 Preliminaries
- The OGR task ranks items for an occasional target group, using user-item and group-item interactions, social connections, and group overlap to address sparse group-item data.The target group’s limited interaction history makes directly learning its preference representation infeasible.Source: §3.1 · p. 6
3.2 Overview of HyperGroup
- HyperGroup combines social-based Individual Preference Modeling (IPM), hyperedge-based Group Representation Learning (HRL), and joint optimization in a hierarchical network.IPM enhances members’ preferences using friends’ preferences before HRL learns group representations from shared-member relationships.Source: §3.2 · pp. 7–8
3.3 Individual Preference Modeling
- IPM uses an inductive social-graph GNN to aggregate users’ own and neighbors’ representations, then averages group members’ embeddings to initialize each group representation.The neighborhood aggregation is mean-based, and the learned aggregation functions generalize to unseen nodes.Source: §3.3 · pp. 8–9
3.4 Hyperedge Embedding-based Group Representation Learning
- HyperGroup models groups as hyperedges and learns their preferences by aggregating information from overlapping groups, rather than representing groups as graph nodes.Its weighted aggregation considers both the number of shared members and the identities of those members, and the method is designed to generalize to unobserved groups.Source: §3.4 · pp. 10–11
3.5 Model Optimization
- HyperGroup trains group-item and user-item recommendation tasks using either two-stage training or joint training.Two-stage training first optimizes user-item interactions and then fine-tunes on group-item interactions; joint training optimizes both losses on their respective interactions.Source: §3.5 · p. 12
- At each SGD step, HyperGroup samples positive user-item and group-item examples, draws N_x corresponding negative samples, and updates gradients for both interaction sets.The positive samples are drawn from the user-item and group-item sets, respectively.Source: §3.5 · p. 13
4 EXPERIMENTAL SETUP
The experiments evaluate HyperGroup on two OGR datasets, using all-item ranking and comparisons with established baselines. They also examine component contributions, optimization, hyperparameters, and scalability.
4.1 Research Qestions
- The experiments ask whether HyperGroup outperforms state-of-the-art methods, how its components and optimization approaches contribute, and how hyperparameters and scalability affect performance.The questions cover IPM, HRL, joint training, and optimization on heterogeneous interaction data.Source: §4.1 · pp. 13–14
4.2 Datasets
- Yelp and Douban-Event provide the evaluation data; their group-item interactions are sparse, and the study excludes datasets lacking suitable occasional groups or social-network information.Yelp averages 1.12 interactions per group versus 20.77 per user; user-item matrix densities are 0.051% for Yelp and 0.057% for Douban-Event.Source: §4.2 · pp. 14–15
4.3 Evaluation Protocols
- Methods are evaluated by ranking all items for each test group, using Top-N Hit Ratio and NDCG after an 80%/10%/10% train-validation-test split.HR counts hits in the Top-N list, while NDCG gives greater weight to hits ranked higher.Source: §4.3 · p. 15
4.4 Baseline Methods
- The baselines span popularity, individual-user methods adapted to groups, topic models, and OGR-specific neural and social approaches.They include Pop, NCF, BPR-MF, PIT, COM, AGREE, SIGR, GroupSA, and GroupIM.Source: §4.4 · pp. 15–16
4.5 Implementation Details
- HyperGroup uses Glorot initialization, Adam, 128-dimensional features and embeddings, dropout of 0.1, and two-stage training by default.IPM uses K=1 layer and HRL K=2; the implementation uses mini-batches of 256 and an initial learning rate of 0.0001.Source: §4.5 · p. 16
5 EXPERIMENTAL RESULTS (RQ1)
HyperGroup significantly outperforms the baselines on both datasets, while its individual-user recommendation performance is also reported as best among the compared methods. Case studies further compare its preference predictions with GroupSA on Yelp.
5 EXPERIMENTAL RESULTS (RQ1)
- HyperGroup significantly outperforms all baselines on both datasets, with all improvements statistically significant at p < 0.01.The comparison is reported in Table 2.Source: §5 · pp. 17–18
- HyperGroup achieves the best performance on the user-item recommendation task among NCF, AGREE, and GroupSA.The authors attribute this to GNN-based individual representations and joint training.Source: §5 · p. 18
- For two randomly chosen Yelp groups, HyperGroup’s predicted scores for ground-truth items are closer to 1 than GroupSA’s.The case study compares preference scores in Fig. 5.Source: §5 · p. 18, Figure 5 · p. 18
6 MODEL ANALYSIS
The section investigates HyperGroup’s components and hyper-parameters, then examines its training efficiency.
6 MODEL ANALYSIS
- The analysis covers component and hyper-parameter ablations, followed by an investigation of HyperGroup’s training efficiency.Source: §6 · p. 19
6.1 Importance of Components (RQ2)
The analysis evaluates the contributions of IPM and HRL, user-item interaction data, and alternative optimization strategies. Variants replace or remove components to isolate their roles.
6.1.1 Importance of IPM.
- HyperGroup significantly outperforms HGroup-SH and HGroup-S on both datasets, supporting the combined use of IPM and HRL.HyperGroup also outperforms HGroup-S, which omits GNN-based individual preference learning.Source: §6.1 · p. 19, §6.1.1 · p. 20
6.1.2 Importance of HRL.
- HyperGroup outperforms HGroup-H, indicating that modeling common-member group similarity with hyperedges helps group representation learning.The reported gap between HyperGroup and HGroup-S is larger than the gap between HyperGroup and HGroup-H.Source: §6.1 · p. 19, §6.1.2 · p. 20
6.1.3 Importance of the User-item Interactions.
- HyperGroup significantly outperforms HGroup-U on both datasets, while HGroup-U outperforms NCF.HGroup-U uses only group-item interactions; its performance gap from HyperGroup is attributed to the extreme sparsity of those interactions.Source: §6.1.3 · p. 20
6.1.4 Comparison of Different Model Optimization Approaches.
- Joint training strategies significantly outperform single-training strategies, and Group-TST performs better than Group-JT.Group-TST first learns individual-user embeddings, then learns group embeddings; Group-JT trains both tasks simultaneously.Source: §6.1.4 · pp. 20–21, Figure 6 · p. 21
6.2 Impact of Hyper-parameters (RQ3)
The section varies sampling, representation, group-size, and item-activity settings, and reports training-loss convergence and feature-initialization findings.
6.2.1 Impact of S
- Sampling more neighbors yields diminishing returns, can reduce recommendation performance beyond a threshold, and increases running time; the chosen S is 4 for both components.Source: §6.2.1 · p. 22
6.2.2 Impact of N_x
- The best-performing negative-sample count is N_x = 4, although experiments use N_x = 1 on both datasets for comparability with.The results show high variance across negative-sample counts.Source: §6.2.1 · p. 22
6.2.3 Impact of w
- The best w is 0.5 on Yelp and 0.3 on Douban-Event; reducing attention to individual preferences does not improve group representations.The authors describe individual preferences as foundations for group preference learning.Source: §6.2.1 · p. 23
6.2.4 Impact of l
- HyperGroup performs best for medium groups (3 ≤ l ≤ 7); small groups offer fewer interactions for similarity modeling, while large groups face greater consensus difficulty.Source: §6.2.1 · p. 23
6.2.5 Impact of τ
- HyperGroup achieves comparable results on cold-start and popular items, with expected performance on popular items.Cold-start items are those with τ ≤ 3; popular items have τ > 3.Source: §6.2.1 · p. 23
6.2.6 Convergence.
- Training loss decreases on both datasets; HyperGroup converges fastest in Yelp’s first 20 iterations and reaches its best result around iteration 30, versus around iteration 10 on Douban-Event.Source: §6.2.6 · p. 23, Figure 7 · p. 23
- Different feature-initialization strategies produce very similar results after HyperGroup is fully trained, so the results are not reported.Source: §6.2.6 · p. 23
6.3 Training Efficiency and Scalability (RQ4)
HyperGroup’s training time grows approximately linearly with training-data scale on Yelp and Douban-Event, providing evidence of scalability.
6.3 Training Efficiency and Scalability (RQ4)
- Yelp training time rises from 0.09×10^3 to 0.535×10^3 seconds, while Douban-Event rises from 0.106×10^3 to 0.617×10^3 seconds as the data ratio increases from 0.2 to 1.0.Source: §6.3 · p. 24
- The authors compare measured training time against an ideal time that scales linearly with the number of training samples.Training time is reported for data ratios of 0.2, 0.4, 0.6, 0.8, and 1.0, with hyperparameters held fixed.Source: §6.3 · p. 24
- Training time grows approximately linearly with data scale on Yelp and Douban-Event, which the authors cite as evidence that HyperGroup scales to large datasets.The experiments vary training-data ratios from 0.2 to 1.0 with hyperparameters fixed.Source: §6.3 · p. 24
7 CONCLUSIONS
HyperGroup learns occasional-group preferences by combining members’ learned individual preferences with group-level hyperedge embeddings. The authors report superior recommendations on two real-world OGR datasets and describe other occasional-group settings where the method can be applied.
7 CONCLUSIONS
- HyperGroup uses a hierarchical GNN and hyperedge embedding to learn group preferences from members’ learned individual preferences.The authors report superiority over state-of-the-art baselines on two real-world OGR datasets.Source: §7 · p. 24
- The approach first uses social interests to address sparse user-item interactions, then models groups as a hypergraph and jointly optimizes user-item and group-item interactions.The hypergraph is used to leverage group similarity in the higher layer of the network.Source: §7 · pp. 24–25
- The authors identify occasional groups at academic conferences and social events as other settings where HyperGroup may apply, including cases without historical group activities.Source: §7 · p. 24