En omfattande utvärderingsram för syntetiska tidsseriedata : Utvärdering av population trohet och användbarhet i samband med klassificering i maskininlärning

In data-driven decision-making, \ac{STSD} addresses challenges like data scarcity, privacy, and the need for robust \ac{ML} models. \ac{TS} data is crucial in fields such as healthcare and finance, but generating high-quality synthetic data that maintains temporal dependencies remains difficult. This work tackles the need for reliable evaluation frameworks to ensure synthetic data fidelity and utility for safer \ac{ML} workflows in sensitive environments. A major challenge is the lack of standardized metrics for assessing \ac{STSD}. Existing methods fail to simultaneously evaluate statistical alignment, distributional similarity, and temporal consistency, leading to inconsistencies in determining reliability. Moreover, the link between fidelity measures and synthetic data usefulness in classification tasks remains unclear, risking misrepresentation and bias in critical \ac{ML} applications. This thesis introduces a framework that integrates \ac{PF} metrics with classification performance. It explores two key questions: (1) How can \ac{STSD} be systematically evaluated across statistical, distributional, and temporal dimensions? (2) How do \ac{PF} metrics correlate with classification utility? The methodology involves data preprocessing, synthetic data generation with quality variations, fidelity evaluation, and statistical correlation analysis. Time series are segmented using sliding windows, and synthetic data is generated across increasing training epochs. Fidelity metrics compare synthetic and real data, while classification models trained on synthetic sets are tested on real holdouts. Spearman’s correlation and confidence intervals assess relationships between PF metrics and F1-ratios, revealing dataset-specific fidelity-utility patterns. The final results highlight dataset-dependent variations. In one dataset, \ac{PF} metrics showed significant correlations with the utility metric at varying strengths, while another dataset lacked notable correlations, reinforcing the context-specific nature of synthetic data evaluation. Future research directions can be applied by expanding \ac{PF} metrics across additional dimensions or assessing datasets from other domains.