A Scalable Machine Learning Generalization Analysis Model for Intelligent Information Systems Incorporating PAC-Bayes Bounds

This study proposes SGAM-PB, a scalable machine-learning generalization analysis model that incorporates PAC-Bayes bounds. The experimental results show that the SGAM-PB achieves the best overall performance across five intelligent information system tasks. Compared with ERM, the generalization gap is reduced from 0.118 to 0.036, a decrease of approximately 69.5%. The PAC-Bayes bound is reduced to approximately 0.15 after 100 training epochs, showing a smoother and tighter convergence trend than the baseline methods. Under 50% drift intensity, the SGAM-PB maintains an accuracy of approximately 0.77, outperforming the ERM by nearly 0.10. In dynamic knowledge update experiments, the recovered accuracy reached approximately 0.86 after 30 update steps, which was higher than that of the ERM and traditional PAC-Bayes baseline. In scalability testing, SGAM-PB maintains the runtime cost below that of the deep ensemble and BNN methods when the data volume increases to 1000 K samples. These results demonstrate that the SGAM-PB can effectively reduce deployment risk, improve drift robustness, accelerate update recovery, and provide interpretable generalization monitoring for intelligent information systems.