SVM Multi-Classification Implementation with Code Enhancements

Support Vector Machines (SVM) are primarily designed for binary classification problems. However, this implementation extends SVM's capabilities to multi-class classification using Chih-Jen Lin's LIBSVM toolkit. The code employs either the one-vs-one or one-vs-rest strategy to decompose the multi-class problem into multiple binary classification subproblems. We validated the approach using the renowned UCI IRIS dataset, which contains 3 classes of iris plants with 4 features each. The implementation involves data normalization, kernel function selection (linear/RBF/sigmoid), and parameter tuning using grid search with cross-validation. Key functions include svmtrain() for model building and svmpredict() for classification, with additional focus on handling class imbalance and optimizing decision boundaries. The results confirm that this multi-class SVM extension effectively handles complex datasets, significantly improving classification accuracy and model stability compared to basic binary approaches. The code demonstrates proper handling of multi-class decision functions and effective hyperparameter optimization for real-world classification scenarios.