Evaluation-oriented data-generating process for performance time series: the perturbation observation decoupling principle

Abstract

This work proposes the Perturbation-Observation Decoupling Principle: by decoupling structured perturbations – defined with precise temporalboundaries and semantic labels – from performance observation paths at the physical, logical, or resource level during the data generation process (DGP), our approach prevents signal contamination and ensures causal traceability. It establishes a controllable, reproducible paradigm for generating performance time series (PTS) with embedded ground-truth annotations, enabling a shift from passive observation to active intervention and providing a causal foundation for high-recall, interpretable, and action-oriented algorithm evaluation.