EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Mathematical Modeling and Bayesian Statistical Learning for Adaptive Artificial Intelligence in Multidimensional Data Analysis

Nouman Shahzad

Journal of Intelligent Decision Making and Information Science · 10.10.2026 · 3 (17s) · 1755-1777

DOI: 10.17762/jidmis.v3.5533 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Intelligent Decision Making and Information Science, online erschienen am 10.10.2026. Verfasst von Nouman Shahzad. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Adaptive artificial intelligence must learn from data whose dimensional structure, predictive relationships, and uncertainty can change after deployment. This paper develops a mathematical and statistical framework, termed the Drift-Aware Bayesian Statistical Learner (DA-BSL), for sequential prediction in multidimensional nonstationary data. The model combines a nonlinear feature map with recursive Gaussian posterior updating and an innovation-sensitive forgetting factor that increases posterior plasticity when standardized predictive surprise is high. A reproducible simulation benchmark was constructed with correlated predictors of dimension 10, 25, 50, and 100; nonlinear response structure; three abrupt concept-drift events; and repeated random seeds. DA-BSL was compared with Bayesian learning without forgetting, fixed-forgetting Bayesian learning, online normalized stochastic gradient descent, and a static ridge model. Across dimensional settings, DA-BSL achieved mean RMSE values of 1.083, 1.131, 1.232, and 1.411, while maintaining 95% predictive-interval coverage near the nominal level at 0.948-0.949. At 100 dimensions, DA-BSL reduced RMSE by approximately 13.9% relative to fixed-forgetting Bayes and 46.1% relative to static ridge, while its negative log predictive density remained substantially lower than the non-adaptive Bayesian baseline. Stress testing showed that prediction error increased with drift severity for all methods, but the proposed approach preserved a more favorable joint profile of accuracy, calibration, and adaptation. An ablation study indicated that standardized uncertainty-aware gating and nonlinear feature construction both contributed materially to performance. These findings support a view of adaptive AI as sequential probabilistic inference rather than repeated point estimation: uncertainty should influence not only what a model predicts but also how rapidly the model is permitted to change. The framework is intentionally transparent and computationally tractable, providing a mathematical baseline for more complex Bayesian neural, continual-learning, and tabular foundation-model systems (Wilson & Izmailov, 2020; Daxberger et al., 2021; Chang et al., 2023; Rudner et al., 2022).

Zitieren

APA

Shahzad, N. (2026). Mathematical Modeling and Bayesian Statistical Learning for Adaptive Artificial Intelligence in Multidimensional Data Analysis. Journal of Intelligent Decision Making and Information Science, 3(17s), 1755-1777. https://doi.org/10.17762/jidmis.v3.5533

Vancouver

Shahzad N. Mathematical Modeling and Bayesian Statistical Learning for Adaptive Artificial Intelligence in Multidimensional Data Analysis. Journal of Intelligent Decision Making and Information Science. 2026;3(17s):1755-1777. doi:10.17762/jidmis.v3.5533

BibTeX
@article{Shahzad2026,
  author = {Shahzad, Nouman},
  title = {Mathematical Modeling and Bayesian Statistical Learning for Adaptive Artificial Intelligence in Multidimensional Data Analysis},
  journal = {Journal of Intelligent Decision Making and Information Science},
  year = {2026},
  volume = {3},
  number = {17s},
  pages = {1755--1777},
  doi = {10.17762/jidmis.v3.5533}
}
Kostenlos starten →