Generalizability of Decomposition Methods

Background: Machine learning models designed for time series forecasting, particularly for complex systems like GPU clusters, often rely on monolithic architectures.

Question / Future Work: The paper describes PRISM’s success in decomposing complex GPU workload signals into interpretable primitives and refining them spectrally. Future work could involve exploring how this primitive decomposition and spectral refinement strategy generalizes to other highly volatile, heterogeneous time series domains beyond GPU cluster workloads, such as financial markets or general IoT sensor data, to assess the broader applicability of the PRISM methodology.