Large-scale pinball twin support vector machines

Twin support vector machines (TWSVMs) have been shown to be effective classifiers for a range of pattern classification tasks. However, the TWSVM formulation suffers from shortcomings: (i) uses hinge loss function which renders it sensitive dataset outliers (noise sensitivity). (ii) It requires matrix inversion calculation in Wolfe-dual is intractable datasets with large numbers features/samples. (iii) minimizes empirical risk instead structural its consequent overfitting. This paper proposes novel scale pinball twin (LPTWSVM) address these shortcomings. The proposed LPTWSVM model firstly utilizes achieve high level noise insensitivity, especially relation data substantial feature noise. Secondly, and most significantly, eliminates need calculate inverse matrices dual problem (which apart being very computationally demanding may not possible due singularity). Further, does employ kernel-generated surfaces non-linear case, using kernel trick directly; this ensures that fully modular approach contrast original TWSVM. Lastly, explicitly minimized improvement accuracy (we analyze properties insensitivity LPTWSVM). Experiments on benchmark show effectively deployed exhibits similar or better performance comparison relevant baseline methods.