Group recommendation aims to recommend tailored items to groups of users, where the key challenge is modeling a consensus that reflects member preferences. Although several deep learning models have improved performance, they still struggle to capture consensus in two important aspects: (1) capturing consensus in small groups (2~5 members), which better reflect real-world scenarios; and (2) balancing individual and group performance while improving overall group accuracy. To address these issues, we propose C$^3$(Capturing Consensus with Contrastive Learning) for group recommendation, which explicitly explores the consensus underlying group decision-making. C$^3$ uses a Transformer encoder to learn both user and group representations, and employs contrastive learning to mitigate overfitting for users with many interactions, resulting in more robust group representations. Experiments on four public datasets show that C$^3$ consistently outperforms state-of-the-art baselines in both user and group recommendation tasks.

翻译：群体推荐旨在向用户群体推荐定制化的项目，其核心挑战在于建模能够反映成员偏好的共识。尽管已有多种深度学习模型提升了性能，但在两个重要方面仍难以有效捕获共识：（1）在小规模群体（2~5名成员）中捕获共识，此类群体更能反映真实场景；（2）在提升整体群体准确率的同时，平衡个体与群体的推荐性能。为解决这些问题，我们提出用于群体推荐的C$^3$（基于对比学习的共识捕获）方法，该方法显式地探索群体决策背后的共识机制。C$^3$采用Transformer编码器学习用户与群体的表征，并利用对比学习缓解高交互频率用户的过拟合问题，从而获得更鲁棒的群体表征。在四个公开数据集上的实验表明，C$^3$在用户推荐与群体推荐任务中均持续优于当前最先进的基线方法。