A unified and scalable machine learning framework for feature fusion in object classification using weighted PCA with adaptive concatenation and dynamic scaling
Amitav Mahapatra, Prashanta Kumar Patra · Discover Computing · 2025
Feature fusion is essential for enhancing the performance of machine learning classifiers, particularly when managing heterogeneous, high-dimensional, and multimodal datasets. In this work, we propose Weighted PCA with Adaptive Concatenation and Dynamic Scaling (WPCA-ACDS) , a novel feature fusion technique designed to address challenges such as overfitting, high dimensionality, and noise sensitivity. WPCA-ACDS integrates three key components: Weighted Principal Component Analysis (Weighted PCA) for efficient dimensionality reduction, Adaptive Concatenation for optimal feature selection based on data-driven strategies, and Dynamic Scaling to balance feature contributions and mitigate the impact of outliers or irrelevant features. Through extensive empirical evaluation on five benchmark datasets—CIFAR-10, Caltech-101, Scene-15, MNIST, and Oxford Pets—utilizing five classifiers (SVM, Random Forest, KNN, Logistic Regression, and Decision Trees), we demonstrate that WPCA-ACDS outperforms several state-of-the-art fusion techniques, including Simple Concatenation , PCA-based Concatenation , Average Fusion , Weighted Average Fusion , Product-based Fusion , cv-weight , Multiple Kernel Learning (MKL) , Collaborative Boosting and Dominant Set Fusion . WPCA-ACDS excels in terms of classification accuracy , robustness to noise and high-dimensional data , and computational scalability . Additionally, sensitivity and trade-off analyses emphasize WPCA-ACDS's versatility, solidifying its position as a robust and scalable solution for modern machine learning tasks. Ablation studies further reveal the critical role of each component, demonstrating that the full WPCA-ACDS framework consistently outperforms all ablated variants in classification performance. This establishes WPCA-ACDS as an effective and comprehensive feature fusion technique for a wide range of machine learning tasks.