【摘要】 现有的序列化推荐算法是假设训练集和测试集来自相同的分布。然而，在实际应用中，用户购买行为的分布可能更为复杂。用户动态变化的偏好、噪声购买行为和新添加的商品都可能导致购买行为分布的变化，使得在训练数据上学习的模型在测试集上无效。为了学习用户的真实偏好，该文构建了一个鲁棒的序列化推荐系统。首先，提出了一种购买行为模拟策略，可以模拟各种购买行为可能出现的复杂情况。其次，建立了一个新颖的投票网络机制，以确保预测结果的稳健性。最后，设计了一个去噪约束来保证学习到的用户偏好表示具有可辨别性。在公开数据集MovieLens-1M、MovieLens-10M和MovieLens-20M上进行验证，并与现有的方法进行比较，在评价指标NDCG@N和Recall@N上较基线方法均有一定提升，验证了该文方法的有效性。更多还原

【Abstract】 Existing sequential recommendation algorithms assume that the train and test sets come from the same distribution. However, users’ preferences variations, noise buying behaviors, and newly added items may all lead to changes in the distribution of purchasing behaviors, making models learned on training data ineffective for practical scenarios. In order to learn the real preference of users, a robust recommender system is constructed. First, a purchase behavioral simulation strategy is proposed to cover the complex situations that may arise in various purchasing behaviors. Second, a novel voting network is built to ensure the robustness of the prediction results based on learned preferences. Finally, a denoising constraint is designed to ensure the learned user preference representation is discriminative. Experimented on the existing methods in the public datasets MovieLens-1M and MovieLens-10M, the proposed method achieves improvements compared with the existing methods in terms of NDCG@N and Recall@N.更多还原