Computational Budget and Scheduler for Graph Structures

This paper introduces the application of a Computational Budget through an adaptive scheduler for graph-based learning architectures to improve computational efficiency. The proposed framework investigates whether a mathematical scheduling strategy can allocate computational resources according to the relative importance of individual graph nodes, thereby reducing unnecessary computation while maintaining predictive performance. The scheduler combines Bayesian uncertainty estimation, Thompson Sampling with an outcome-based reward structure, and a new composite metric, Meditation (MED), while Computational Budget serves as the primary bookkeeping measure of resource allocation. The framework was evaluated on both synthetic graph structures and the real-world Cora and Citeseer citation network datasets. Experimental results demonstrate that adaptive scheduling can reduce computational cost and runtime while maintaining accuracy within an acceptable margin under favorable graph conditions, highlighting the potential of adaptive resource allocation as a direction for more efficient graph learning.