The authors propose Temporal-Decay Shapley methods for valuing training samples in time-series data, addressing the fact that standard Shapley-value methods assume i.i.d. data and ignore the time-varying importance of samples. Their best method, MS-TDS, uses multiple exponential-decay timescales in parallel and fuses them adaptively per sample, balancing short-term "hotspot" data and long-term foundational data. Experiments show the approach outperforms traditional Shapley methods for noise detection and high-value sample identification, especially when temporal drift is strong.
Main takeaways:
- Standard Shapley data valuation assumes samples are independent; this breaks for time-series data where recency and temporal drift matter.
- Temporal-Decay Shapley (TDS) weights samples by exponential decay; improved TDS uses power-exponential decay for nonlinear drift.
- Multi-Scale TDS (MS-TDS) runs parallel decay scales and fuses them per sample, capturing both recent trends and long-term patterns.
- Empirically, the temporal methods beat baselines on noise detection and data-selection tasks, with larger gains under strong temporal settings.