Adaptive schedule deviation correction for complex information system projects: a graph neural network approach with knowledge distillation

Schedule deviation continues to plague complex information system projects, with empirical surveys reporting that over 70% of such projects suffer meaningful overruns and average delays exceeding one-quarter of the originally planned duration. Conventional schedule management tools—earned value analysis, critical path computation, and linear regression—rest on assumptions of stable task environments and independent activity progression, assumptions that rarely hold when hundreds of interdependent tasks compete for shared resources under volatile requirements. More recent machine-learning predictors, including LSTM-based sequence models and random forests, improve point-wise accuracy yet still treat tasks in isolation or impose chain-like orderings, discarding the rich graph topology that governs real project execution. Equally problematic, deep predictive models often demand computational budgets that preclude the real-time responsiveness project managers need on the ground. This paper introduces an adaptive correction framework that addresses both shortcomings simultaneously. We model task dependencies as a graph processed by a hierarchical attention network, then compress the learned representations into a lightweight student model through a dual-tower knowledge distillation strategy, and finally close the loop with a reinforcement-learning module that selects corrective actions in response to predicted deviations. Experiments on 234 real-world projects show that the framework achieves an RMSE of 4.87% and an R² of 0.847, outperforming established baselines, while the distilled student model completes inference within a few milliseconds—fast enough for continuous schedule monitoring even on resource-constrained hardware.