Performance Enhancement of Multi-Class SVMs using Symmetric ADMM: Node-based Scalability for Large- Scale Classification

Along with the fast development of large-scale and multiclass datasets, there is a requirement for efficient optimization frameworks for Support Vector Machines. Highperformance identification is achieved by conventional formulations like the Crammer-Singer (CS-SVM) and Weston-Watkins (WW-SVM), but they suffer from high computational overhead when training on massive datasets. We propose a distributed optimizational framework that couples S-ADMM with CS-SVM and WW-SVM for an incremental dataset and incremental computation nodes in this article. Our method improves scalability by dividing the global optimization problem into parallel sub-problems while symmetric updates guarantee speedy convergence and reduced communication delay. The experiments on the benchmark datasets such as LSHTC demonstrate significant improvements in training time and convergence stability compared to the traditional methods. The results indicate that the proposed framework, in fact, preserves the classification accuracy and obtains up to 30% reduction of training time and 4.5 times time increase in speedup with incrementally scaled nodes, thus, it can be considered as a reliable solution in large-scale, distributed machine learning environments.