TimeLAVA: Learning-Agnostic Valuation for Time Series Data

Data valuation quantifies the intrinsic quality of individual samples to enable principled data curation, quality control, and robust learning. For time series in critical domains such as healthcare, finance, and industrial monitoring, effective valuation methods are essential yet fundamentally lacking. Existing approaches are either model-dependent, limiting their generalizability, or designed for i.i.d. data and thus fail to capture temporal dependencies, multi-scale patterns, and non-stationary dynamics inherent to sequential data. We introduce TimeLAVA, a learning-agnostic framework that values temporal segments by their marginal contribution to minimizing distributional discrepancy between evaluated and reference data. At its core is a novel Selective Wavelet-based Wasserstein ($\mathcal{W}_\text{SW}$) discrepancy combining multi-scale wavelet transforms for temporal localization with unbalanced optimal transport for robustness to distributional shifts. Segment values are efficiently computed via sensitivity analysis without requiring model training and aggregated into point-wise scores. We 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.

In many critical applications such as hospital patient monitoring, financial trading, and industrial equipment maintenance, decisions rely on continuous streams of data collected over time. However, not all data is equally useful: some portions may be corrupted by sensor faults, contain rare abnormal events, or be mislabeled by human annotators. Training machine learning models on such low-quality data can lead to unreliable or even dangerous predictions. We address a fundamental question: how can we automatically score each piece of time series data based on its intrinsic quality, without needing to first train a specific prediction model? This is challenging because time series data has unique properties. Observations depend on what came before, patterns occur at different time scales, and statistical properties can change over time. We introduce TimeLAVA, a framework that assigns quality scores to segments of time series data by measuring how well each segment matches a trusted reference dataset. Our approach combines two key ideas: (1) a multi-scale representation that captures both short-lived spikes and long-term trends in the data, and (2) a flexible matching strategy that allows genuinely unusual segments to remain unmatched rather than being forced into misleading comparisons. Importantly, the resulting scores are computed efficiently without training any predictive model, making them applicable across diverse tasks. We provide mathematical guarantees that our scores are stable in the presence of data corruption and are meaningfully connected to downstream model performance. Experiments across anomaly detection, data cleaning, and mislabel identification on real-world healthcare, finance, and industrial datasets show that TimeLAVA produces substantially more informative quality assessments than existing methods, enabling practitioners to curate their data more effectively before model training.