基于用户协同过滤的电商多场景推荐方法研究

Research on Multi-Scenario Recommendation Method for E-Commerce Based on User Collaborative Filtering

摘要: 随着电商平台多元化业务形态的发展，用户行为呈现出跨场景分布特征，传统单场景推荐方法难以有效刻画用户复杂偏好。为解决电商多场景推荐中随机负采样难以刻画真实负反馈、传统贝叶斯个性化排名损失对样本难度区分不足以及跨场景知识迁移受限等问题，提出一种基于用户协同过滤与自适应BPR损失的多场景推荐方法(EDDA-UBCF)。该方法在EDDA框架基础上，构建融合用户历史行为与相似用户偏好模式的协同过滤负样本采样策略，以提高负样本质量；设计引入动态边际系数、样本权重和自适应正则化的增强型BPR损失，以提升模型对困难样本及场景差异的建模能力；同时结合嵌入解纠缠与场景对齐机制，强化跨场景共享知识与场景特定知识的协同学习。基于支付宝APP五场景真实电商数据集AntM2C开展实验，并与11种基线模型进行比较。结果表明，所提方法在多个子场景取得最优或次优性能。消融实验进一步验证了协同过滤负采样策略与自适应BPR损失的有效性。该方法能够更准确地捕捉用户跨场景偏好迁移特征，提升多场景推荐的准确性与鲁棒性，可为电商平台个性化推荐提供有效支撑。

Abstract: With the development of diversified business formats on e-commerce platforms, user behaviors have exhibited cross-scenario distribution characteristics, making it difficult for traditional single-scenario recommendation methods to effectively capture complex user preferences. To address the problems in e-commerce multi-scenario recommendation that random negative sampling fails to characterize real negative feedback, the traditional Bayesian Personalized Ranking (BPR) loss lacks the ability to distinguish sample difficulty, and cross-scenario knowledge transfer is limited, this paper proposes a multi-scenario recommendation method based on user-based collaborative filtering and adaptive BPR loss, termed EDDA-UBCF. Built upon the EDDA framework, the proposed method constructs a collaborative-filtering-based negative sampling strategy by integrating users’ historical behaviors with preference patterns of similar users to improve negative sample quality; it further designs an enhanced BPR loss incorporating dynamic margin coefficients, sample weights, and adaptive regularization to strengthen the model’s ability to learn from hard samples and scenario differences. Meanwhile, by combining embedding disentanglement and scenario alignment mechanisms, the method enhances the collaborative learning of cross-scenario shared knowledge and scenario-specific knowledge. Experiments are conducted on the real-world five-scenario e-commerce dataset AntM2C from the Alipay App and compared with 11 baseline models. The results show that the proposed method achieves the best or second-best performance in multiple sub-scenarios. Ablation studies further verify the effectiveness of the collaborative-filtering-based negative sampling strategy and the adaptive BPR loss. The proposed method can more accurately capture users’ cross-scenario preference transfer characteristics, improve the accuracy and robustness of multi-scenario recommendation, and provide effective support for personalized recommendation on e-commerce platforms.