Label noise and irregular sampling often co-occur in real-world sensor and log streams, degrading model reliability. Our proposed **Noise-Aware Risk Estimation (NARE)** framework: * **Models timestamp-level noise** via a Beta-Bernoulli prior, enabling adaptive confidence weighting. * **Derives a closed-form risk estimator** that upper-bounds the true empirical risk under irregular sampling. * **Incorporates a co-teaching curriculum** that gradually filters high-risk instances during training. On three public IoT and healthcare benchmarks, NARE improves macro-F1 by up to **8 pp** over state-of-the-art robust TSC baselines, while offering calibrated uncertainty scores for critical-event detection.Recommended citation: Kun Han, Abigail M.Y. Koay, Ryan K.L. Ko, Weitong Chen & Miao Xu (2025). “Mining Irregular Time Series Data with Noisy Labels: A Risk Estimation Approach.” *Proceedings of the Australasian Database Conference (ADC 2025)*, pp. 293–307. Springer, Singapore.
Download Paper | Download Slides
