TimeLAVA: New method values time series data without model training

Data valuation — assigning a quality score to each data point — is critical for curating high-quality datasets, detecting label noise, and pruning redundant samples. For time series data in domains like healthcare, finance, and industrial monitoring, existing valuation methods either rely on training models (limiting generalizability) or assume independent and identically distributed (i.i.d.) data, failing to capture temporal dependencies, multi-scale patterns, and non-stationary dynamics. A new paper from researchers at the University of Melbourne, the University of Adelaide, and Amazon introduces TimeLAVA, a learning-agnostic framework specifically designed to value time series segments without requiring model training.

At its core, TimeLAVA measures the marginal contribution of a temporal segment to minimizing the distributional discrepancy between the evaluated dataset and a reference dataset. It uses a novel Selective Wavelet-based Wasserstein discrepancy that combines multi-scale wavelet transforms for temporal localization with unbalanced optimal transport to handle distributional shifts robustly. The segment values are efficiently computed via sensitivity analysis, then aggregated into point-wise scores. The authors provide theoretical guarantees linking valuation to model-agnostic generalization and prove bounded sensitivity to outlier contamination. Extensive experiments across anomaly detection, data pruning, and label noise detection demonstrate that TimeLAVA produces significantly more informative value scores than existing methods on diverse real-world datasets.