Abstract: Click-through rate prediction is a critical task in recommendation systems, serving as the final step in recommending items to users. Most existing state-of-the-art methods focus on modeling complex implicit and explicit feature interactions; however, they often overlook spurious correlations caused by confounding factors, which reduces the model's generalization ability. To address these challenges, A Federated Click-through Rate Prediction Framework Based on Relevance Screening (FedCRP) is proposed, comprising two key components. First, the Dynamic Recurrent Collaboration Network (DRCN) enables the model to capture various nonlinear couplings from multi-scale feature structures. Second, the Relevance Screening Strategy (RSS) employs variable isolation analysis and sample priority adjustment to eliminate spurious correlations in complex feature interactions, allowing the model to focus on causal features. Finally, the model is trained using a federated learning approach, leveraging diverse feature interactions to avoid reliance on single-source spurious correlations. Experiments conduct on four public datasets demonstrate the feasibility and effectiveness of FedCRP. Compared to the best baseline, the evaluation metric AUC improves by 4.10%, 5.29%, 1.48%, and 0.59% across the four datasets, respectively.

摘要： 点击率预测是推荐系统中的一项关键任务，是为用户进行项目推荐的最终步骤。现有的大多数前沿方法主要集中于研究复杂的隐式和显式特征交互；然而，这些方法忽略了混杂因素引起的虚假相关问题，从而降低了模型的泛化能力。为了解决上述挑战，提出了一种基于相关性筛选的联邦点击率预测框架(Federated Click-through Rate Prediction, FedCRP)，它包含两个关键组件。首先，动态循环协作网络(Dynamic Recurrent Collaboration Network, DRCN)使模型能够从多尺度特征结构中挖掘各类非线性耦合。其次，相关性筛选策略(Relevance Screening Strategy, RSS) 利用变量隔离分析和样本优先级调节消除复杂特征交互中的虚假相关，使模型专注于触发特征。最后，使用联邦学习方法训练模型，利用多样化的特征交互，以避免对虚假相关的单一依赖。在4个公开数据集上进行实验，实验结果表明了FedCRP的可行性和有效性。相较于最优基线，评估指标AUC在4个数据集上的提升分别为4.10%、5.29%、1.48%和0.59%。