AN ITERATIVE MODEL USING QROBL-SCA AND MO-CI-GAT FOR OPTIMIZATION AND SCALABLE FEATURE SELECTION IN BIG DATA ENVIRONMENTS

Abstract

The phenomenal increase on high dimensions and big datasets in modern computational environments requires sophisticated efficient optimization techniques, scalable feature selection, and high accuracy classification methods. The traditional algorithms such as the standard Sine Cosine Algorithm (SCA) feature selection and others face slow convergence, premature spot-hold in local optima that result to inefficiency when handling redundant or irrelevant features specially in big data settings. Hence, this work proposes a completely integrated and distributed hybrid framework combining five analytic novel methods suitable for optimization, feature selection, and classification under complex data conditions. The first one is through the Quantum Rotational Opposition-Based Learning Sine Cosine Algorithm (QROBL-SCA) which improves speed of global convergence and search diversity among global alternatives through quantum Inspired phase rotations and opposition-based reinitialization. Second, the Multi-Objective Chaos Induced Graph Attention Network (MO-CI-GAT) captures inter-feature relations as well as maximizes inter-classability by chaotic initializing of attention weights and including Pareto-driven objectives. Thirdly, a Federated Swarm Feature Alignment using Adaptive Mutual Information (FSFA-AMI) provides feature selection across distributed nodes which is consistent globally, while maintaining privacy. Fourth, the Temporal-Spatial Ensemble Support Vector Machine with Dynamic Kernel Fusion (TS-ESVM-DKF) embraces temporal and spatial dependencies through adaptive kernel learning for onward dynamic classification of datasets. Finally, the Reinforcement Learning-Driven Data Partitioning and Feature Batching in MapReduce (RL-DPFB-MR) maximize resource allocation and effectiveness of execution in distributed environments through a deep Q-learning agent. A proposed architecture gives 30-50% improvements in features reductions, 2-3 speedup in distributed classification, and up to 97% accuracy increase on benchmark datasets & samplings. This study expands the intelligent scalable learning pipelines for real-time decision systems and high-dimensional data samples.