Weight Balanced Linear Programming SVM on Skewed Distribution and its Evaluation

Support Vector Machines (SVM) have been shown to have strong classification capability and good generalization results. Among SVM´s, linear-programming SVM´s (LPSVM) have been shown to have better ability to adapt to large data sets with more efficiency. However, there have been several variations of LPSVM´s, and their ability to tackle with skewed distributions are not investigated yet. In this research, we compared 4 LPSVM formulations and suggested one of them as the candidate for classifying skewed distributions. We also suggested a novel formula for evaluation of correctness of classification results, which is more suitable in multicategory cases. We also introduced a new concept of output balancing, which can be useful in recovering lost classes in k-SVM applications. Through the experiments, we showed that weight balancing approach is effective in multicategory classification.