Abstract: With increasing applications of personalized recommendation systems on a variety of platforms, the need to accurately understand and model complex user behaviors along with vast item information has become critical. Traditional recommendation systems often overlook difficult samples that arise from user curiosity or misoperations. These challenging samples, if not properly addressed, can lead to model bias and a significant drop in performance. Additionally, traditional recommendation systems often only consider individual user-item interactions, failing to capture higher-order relations from the perspective of distributions. To address the above limitations, this paper proposes a novel distribution-aware optimization for robust recommendations (DORRec), which targets the matching between global user and item distributions without supervision while distinguishing difficult samples to achieve robust recommendations. Specifically, in the distribution-based difficult sample distinguishing module, this paper relaxes the regularization constraints under the Sinkhorn distance to compute closed-form solutions for complex user-item matching scores, so as to find the hard samples for each user. Furthermore, in the adaptive threshold-based high-robustness recommendation module, this paper proposes a personalized threshold mechanism that adaptively adjusts interaction weights to enhance the training of difficult samples, meeting the need for robust recommendations. Experiments on four public datasets validate the effectiveness of the DORRec framework, demonstrating significant improvements in accuracy and robustness. Compared with multiple state-of-the-art recommendation algorithms and components on several evaluation metrics, DORRec demonstrates significant superiority in recommendation performance.

Key words: recommendation system, difficult samples, distribution-aware, robust recommendations, optimal transport

摘要： 随着个性化推荐系统在各类平台上的广泛应用，如何精确理解并建模复杂的用户行为和海量物品信息成为关键挑战。传统推荐系统经常忽视由好奇心或误操作产生的困难样本，这些未经处理的困难样本如果处理不当，则可能会导致模型偏差和性能降低。此外，传统推荐系统往往仅考虑单一用户-物品交互，未能从分布的角度捕捉高阶关联。针对上述问题，提出了一种基于分布感知优化的高鲁棒推荐方法（DORRec），旨在无监督条件下匹配全局用户分布与物品分布，同时甄别困难样本并建模分布匹配，以实现高鲁棒性推荐。在基于分布的困难样本识别模块，通过利用放宽Sinkhorn距离下的正则化约束，计算复杂的用户-物品间匹配分数的闭式解，从而找到每个用户的困难样本；在基于自适应阈值的高鲁棒推荐模块，提出了一种个性化阈值机制，通过自适应调整交互权重以强化困难样本的训练，满足高鲁棒推荐需求。在四个公共数据集上的实验验证了该方法的有效性，显示出DORRec在准确性和鲁棒性上的提升，通过与多个最先进推荐算法和高鲁棒组件在多个评价指标上进行比较分析，验证了DORRec在推荐性能上的显著优越性。

关键词: 推荐系统, 困难样本, 分布感知, 鲁棒推荐, 最优传输