CORF-Net: Cross-Order Representation Fusion Network for Denoised Survey Response Modeling

Abstract

Improving recommendation quality and user satisfaction is critical to the success of recommendation platforms. Prior studies have attempted to enhance user experience by collecting user feedback through surveys and modeling their responses accordingly. However, the inherent noise in survey responses presents a significant challenge, limiting both the effectiveness of model training and the accuracy of the recommendation ranking. Effectively identifying noise in survey responses and modeling denoised samples requires an expressive and robust model architecture. Unfortunately, most existing recommendation models overlook the importance of capturing high-order feature interactions and their synergistic fusion with low-order features for complementary learning. To address these challenges, we propose a novel Cross-Order Representation Fusion Network (CORF-Net) and a corresponding two-phase denoising strategy. CORF-Net is architecturally designed to capture and integrate both high- and low-order feature interactions, enhancing representation learning from complex survey data. Our two-phase strategy first filters out low-quality samples during training to ensure model robustness, and then dynamically re-weights survey signals based on predicted response quality during ranking to optimize recommendations. Offline experiments demonstrate the architectural superiority of CORF-Net, which achieves up to a 0.92% relative improvement in AUC and a 1.22% reduction in LogLoss over state-of-the-art models. Furthermore, a 45-day large-scale online A/B test on TikTok confirms its real-world impact, yielding a 0.025% increase in user active days, a 0.255% rise in ''Like'' interactions, and significant reductions in negative feedback, including ''Skip'' (-0.418%), ''Report'' (-3.214%), and ''Dislike'' (-2.041%).