A functional iterative approach for twin bounded support vector machine with squared pinball loss (Spin-FITBSVM)

Twin support vector machine (TSVM) plays a significant role in strengthening the generalization performance in the area of binary classification by considering a couple of smaller-sized quadratic programming problems (QPPs). It takes significantly lower learning cost in contrast to support vector machine (SVM). However, it is less stable and sensitive towards noise, like SVM, which is one of the drawbacks that motivates making an algorithm more robust. To alleviate the mentioned demerit, in this work, we propose a new functional iterative approach for twin-bound SVM with squared pinball loss (Spin-FITBSVM). This approach has the following advantages, i.e., more robust, strongly convex and provides strong stability for resampling. To reduce the time complexity, the solution is obtained by using a functional iterative approach instead of a pair of dual quadratic programming problems solved in TSVM. So, it does not have any significant need for any external optimization toolbox while attaining the solution. The numerical experiments have been employed on standard publicly available as well as artificial datasets to validate the fruitfulness and superiority of the proposed Spin-FITBSVM. The outcomes are compared with baseline and recent models like SVM, TSVM, TSVM with pinball loss (PL) function (pin-TSVM), general TSVM with PL function (pin-GTSVM), generalized Huber twin SVM (GHTSVM) and sparse pinball twin SVM (SPTWSVM) for noisy corrupted datasets, which reveals the applicability of the proposed Spin-FITBSVM.