Authors: Deepshikha Rajpoot, Sonam Chamber, Poonam Bhartiya
Abstract: With the evolution of digital transformation landscape, today data stream analytics has become an essential paradigm involving real-time decision-making in various domains such as finance, healthcare and Internet of Things (IoT). Nonetheless, the fact that data streams are non-stationary–concept drift, noise or change pattern are some of the difficulties for traditional static ma-chine learning models. In this research, we provided a comprehensive meta-analysis of the existing self-learning AI architec-tures to analyze adaptive data streams. This study compares the performance of these models using findings from recent litera-ture across three key criteria: robustness to environmental change, scalability in high-throughput scenarios and prediction accu-racy over longer time horizons. This meta-analysis combines information from 30+ landmark studies by evaluating Adaptive Random Forests, Evolving Fuzzy Systems, and Self-Adjusting Memory (SAM) models. The experiments show that deep learning-based adaptive systems achieve better predictive reliability, while we found that some instance-based learning models (i.e., SAM-kNN) perform well on various heterogeneous drift types without retraining effort and hyperparameter tuning. In addition, we discover a general trade-off between computational scalability and adaptivity granularity. The approach used is that of a systematic review, with quantifiable measures to evaluate architectural performance. In this paper, we further discuss the impact of delayed labeling and noise interference on the stability of adaptive learners. We conclude this paper with the implica-tions of these findings in building explainable and energy-efficient AI systems, along with a few future research directions merging federated stream learning with generative AI. This review provides a practical starting point for researchers and practi-tioners that aim to deploy resilient, autonomous and scalable AI systems in dynamic and high-throughput data environments by bridging the gap between theoretical frameworks and deployment challenges.
