Dual-Challenge Framework for Noise-Resilient and Temporally-Adaptive Recommender Systems

Abstract

Collaborative filtering recommender systems are a cornerstone of personalized recommendation; however, they typically assume that explicit user feedback is reliable and that item relevance remains static over time. In practice, user–item ratings often contain noise due to inconsistent user behavior, and item preferences evolve with temporal context. Existing methods generally address only one of these challenges, either filtering out noisy ratings or adapting to temporal drift, resulting in suboptimal performance when both factors coexist, as in real-world scenarios. To overcome this limitation, this paper introduces RADAE (Residual Attentive Denoising Auto-Encoder), a dual-challenge framework that jointly performs noise resilience and temporal adaptation for collaborative filtering. RADAE embeds users, items, timestamps, and corrupted ratings into continuous latent spaces to preserve semantic meaning and interaction chronology. The encoder employs stacked Conv1D layers to capture local temporal dynamics, residual connections for stable learning, and self-attention to emphasize reliable behavioral patterns. The decoder reconstructs denoised, time-consistent ratings through dense expansion and transposed convolutions. A Huber loss objective function further balances fine-grained accuracy with robustness to outliers. Experiments on three benchmark datasets demonstrate that RADAE outperforms classical and recent baselines, significantly improving collaborative filtering performance by producing adaptive, denoised rating predictions aligned with evolving user preferences and item relevance.