Source-linked AI summary
A Flexible Recommendation System for Individuals and Groups
Yacine Mokhtari, Grégory Smits
TL;DR
Group recommenders either aggregate individual preferences or learn representations from group histories, while overlooking preference adaptation or facing limited group data. This paper learns individual and group-context preferences and uses user profiles to guide group aggregation. Experiments report recommendations for individuals and seen and unseen groups, with different comparative results across settings.
Problem
OGRs assume preferences stay stable in groups, while PGRs rely on group histories and cannot generalize to unseen or rarely observed groups.
Source: §1 · p. 2Method
A graph-based approach learns individual and group-context user representations, profiles adaptability using preference shifts and item diversity, and applies group aggregation strategies.
Source: §2.2.1 · p. 3, §2.3 · pp. 4–5, §2.3.1 · p. 5Results
The approach outperforms baselines across top-20 individual recommendation metrics and OGRs for unseen groups, while KGAT-PGR is slightly better overall for seen groups with abundant interactions.
Source: §3.2 · p. 8Takeaways & Limitations
Experiments report relevant recommendations for individuals, seen groups, and unseen groups, prioritizing non-adaptive members’ preferences without discarding adaptive members.
Source: §4 · p. 8Experiments with real datasets and possibly synthetic knowledge graphs, along with robustness analysis and ablation studies, are still being conducted.
Source: §4 · p. 8
Abstract
from arXiv · showhide
Group recommender systems typically rely on either aggregating individual preferences or treating groups as distinct meta-users. However, these methods often suffer from static aggregation strategies or data sparsity issues within group histories. This paper introduces a novel approach, that relies on a GNN-based architecture to learn a dual representation of each user's preferences, capturing their behavior as an independent individual from one side and as a member of a collective from the other side. By performing a differential analysis of these individual and group-oriented preferences, our system then determines the behavioral profile of each user when joining a group. Finally, specific preference aggregation strategies are defined to cope with the behavioral profiles of the users composing a group. Consequently, the system is equally capable of delivering precise recommendations to individuals and to arbitrary groups, effectively unifying the two traditional paradigms of recommendation. Experiments on synthetic data simulating diverse group settings and behaviors confirm the flexibility and relevance of the proposed approach compared to state-of-the-art methods.
1 Related Work
Group recommendation methods divide into OGRs, which aggregate individual preferences, and PGRs, which learn representations for group compositions. Their limitations include static preference assumptions in OGRs and sparse group histories that constrain PGRs.
1 Related Work
- OGRs aggregate members’ individual preferences or recommendations, generally using average, minimum, or maximum satisfaction, without leveraging past group interactions.They assume users’ preferences remain stable when joining groups.Source: §1 · p. 2
- Knowledge graphs have been reported to improve recommendation in individual and group settings, and KGAG is identified as the state-of-the-art KG-based group recommender on simulated data.Source: §1 · p. 2
- PGRs learn representations for each group composition from past interactions, but limited group data restricts generalization to unseen or rarely observed groups.Examples include AGREE, SoAGREE, GroupSA, GroupIM [17], ConsRec [22], and SGGCF.Source: §1 · p. 2
2 Flexible Recommendation System
The system learns individual and group-member preferences, profiles how users adapt in groups, and uses those profiles to tailor group recommendation aggregation. Its components combine graph-based preference learning with interpretable, profile-aware recommendation strategies.
2.1 Context and Notations
- Individual and group interactions are represented separately, allowing the system to infer each user's preferences in both contexts.Source: §2.1 · p. 3
2.2.1 KGE layer
- TransR initializes KG entity and relation embeddings, while a user-specific modulation vector transforms the user embedding for group interactions.A contrastive loss trains on positive KG triplets and corrupted-tail negatives, incorporating domain-KG semantics into the initial embeddings.Source: §2.2.1 · p. 3
2.2.2 GAT layers
- GAT layers aggregate KG neighborhood information into final embeddings, and BPR optimizes matching scores for individual and group-member interactions.The two interaction training sets include valid interactions and randomly sampled negative examples.Source: §2.2.2 · pp. 3–4
2.3 Recommending to Groups
- The system profiles adaptability through preference duality and group-item diversity, then classifies users as adaptable, constant, or dual.The classification rules use adjustable duality and diversity thresholds; diversity is average pairwise dissimilarity among items interacted with in groups.Source: §2.3 · pp. 4–5
2.3.1 Aggregation Strategy wrt. Group Configuration
- For different group compositions, a profile-aware bipolar strategy uses least-misery preferences to select candidates and average preferences to rerank them.The strategy prioritizes non-adaptable users while also seeking to satisfy adaptable members; its rationale for dual users depends partly on whether their group preferences align with constant users or reflect sparse group context.Source: §2.3.1 · p. 5
3 Experiments
Experiments compare recommendations for individuals and seen and unseen groups using generated datasets designed to vary group histories, compositions, and user behaviors. The proposed method performs consistently across these settings, with its relative performance depending on whether group interaction histories are available.
Datasets
- Generated datasets reproduce behaviors and group compositions observed in CAMRa2011 and Mafengwo, representing low-diversity/high-duality and high-diversity/low-duality settings, respectively.Adaptive users were also generated to represent additional user behavior.Source: Datasets · p. 6
- The datasets vary group histories: some groups have few item interactions to challenge PGRs, while genDatalarge has extensive histories to assess generalization.genDatasmall 2 also gives users larger individual histories than genDatasmall.Source: Datasets · p. 6
Experimental Settings
- Evaluation uses an 80/20 individual-interaction split and withholds 10% of groups from training; recommendations are assessed with HR and NDCG at K = 20.The remaining groups are split 70/30 for training and testing.Source: Datasets · p. 6, Experimental Settings · p. 6
3.2 Performance Evaluation
- The approach consistently outperforms all baselines for individual recommendations across top-20 evaluation metrics.The authors attribute this to modeling two user representations, which enhances collaborative signals.Source: §3.2 · p. 8
- For seen groups with many interactions, KGAT-PGR remains slightly better overall, while the proposed method is competitive without learning a separate representation for each group.It aggregates group interactions per user and applies user profiling and adaptive aggregation based on group composition.Source: §3.2 · p. 8
- For unseen groups, the approach generalizes well, gives more accurate recommendations, and outperforms OGR models; PGRs cannot recommend groups absent from training.Source: §3.2 · p. 8
4 Conclusion & Perspectives
The approach learns individual and group-context preference representations, compares them to profile user adaptability, and prioritizes less-adaptive members without discarding adaptive members. The authors report flexibility across individuals and seen and unseen groups, while noting that real-data and robustness evaluations remain underway.
4 Conclusion & Perspectives
- The method learns individual and group-context preference representations, compares them to determine adaptability, and prioritizes non-adaptive members while retaining adaptive members' preferences.Source: §4 · p. 8
- Experiments across varied group settings report relevant recommendations for individuals and both seen and unseen groups.Source: §4 · p. 8
- Evaluation on real datasets and robustness analysis across hyperparameter settings are still underway.The authors also plan an ablation study of the approach's components.Source: §4 · p. 8