Learning TWSVM using Privilege Information

Abstract

Expert’s knowledge can be used to improve classification performance of the algorithm or to reduce the requirement of data for training the algorithm. However, in the field of machine learning, the knowledge offered by the expert is rarely used. Recently, Qi et al. [1] proposed a fast learning model for TWSVM using privilege information termed as FTWSVMPI where privilege information is acquired by Oracle function. Oracle function needs to solve two additional TWSVM based Quadratic Programming Problems (QPPs) which leads to higher computational cost. Therefore, to avoid to solve two additional TWSVM based QPPs, in this paper, we propose a novel method to extract privilege information from the dataset itself. Using this privilege information, we further introduce an improved version of Twin Support Vector Machine termed as I-TWSVMPI. The proposed I-TWSVMPI incorporates privilege information using correcting function so as to obtain two nonparallel hyperplanes. We also perform experiments for pedestrian detection as an application to proposed I-TWSVMPI. The experimental results on several benchmark UCI datasets and pedestrian detection prove the efficacy of our proposed formulation to that of other state-of-the-art classification algorithms with comparatively lesser computational time.