RoFeR: Robust Feature Enhancement and Representation Optimization for Noisy Time-Series Data

Noisy environments pose significant challenges to the analysis of time-series data, leading to degraded performance in downstream tasks such as prediction, classification, and anomaly detection. In this paper, we propose RoFeR, a novel framework for Robust Feature Enhancement and Representation Optimization of time-series data under noisy conditions. RoFeR integrates multi-scale adaptive feature enhancement with representation learning to capture essential temporal patterns while suppressing noiseinduced distortions. Specifically, it employs a hybrid mechanism combining attentionbased filtering and embedding optimization, ensuring that the learned representations are both informative and resilient to noise. The core contributions of RoFeR include (1) a multi-scale convolutional feature extraction approach that effectively captures both short-term and long-term temporal dependencies, (2) an adaptive attention mechanism for noise-aware feature fusion, and (3) a robustness-driven embedding optimization step to ensure the stability of the learned representations under various types of noise. These steps make RoFeR different from existing methods by seamlessly integrating feature enhancement with noise-resistant learning, enabling better performance in noisy environments. Extensive experiments on benchmark datasets demonstrate that RoFeR significantly improves robustness and accuracy compared with state-of-the-art baselines, highlighting its potential in real-world applications involving noisy temporal signals.