TimeLAVA: Learning-Agnostic Data Valuation for Time Series

Researchers have introduced TimeLAVA, a learning-agnostic framework for valuing time series data based on its marginal contribution to minimizing distributional discrepancy.

TimeLAVA is designed for critical domains such as healthcare, finance, and industrial monitoring, where existing data valuation methods are either model-dependent or fail to capture temporal dependencies^[1]. The framework uses a novel Selective Wavelet-based Wasserstein discrepancy for robustness to distributional shifts. Segment values are computed via sensitivity analysis without model training. In related research, model selection between probabilistic models on time series data sets is considered using proper scoring rules. Three summary statistics are commonly used for model selection: mean score, median score, and mean rank. The mean score is effective for short test sets, while model selection based on mean ranks remains stable with different scaling factors^[2]. However, the distribution of scores can be skewed, leading to conflicting decisions, highlighting the importance of a large test set for accurate model selection.