Robust nonparallel support vector machine with privileged information for pattern recognition

In the field of pattern recognition, collected data always include some additional information which are usually termed as privileged information. The privileged information is latent information belonging to the training samples, which can be easily ignored. The privileged information can help to build a better classifier for classification. In this paper, we try to construct a robust nonparallel support vector machine (NPSVM) model under the privileged information learning (LUPI) setting, termed as R-NPSVM+. On the one hand, we introduce the privileged information into NPSVM so as to build a model for classification. In the training process, both the privileged information and usual samples are used to train the model, which can enhance the accuracy. On the other hand, due to the epsilon-insensitive loss and hinge loss, NPSVM is sensitive to noise or outliers. Hence, we use two robust loss functions in R-NPSVM+ model, which can further ensure the robustness of the model. In addition, we use the Lagrange multiplier method and the dual coordinate descent (DCD) algorithm to optimize the proposed objective function, respectively. Lastly, to evaluate the performance of R-NPSVM+, we conduct a series of experiments. Experimental results confirm that compared with other classical SVM-type algorithms, our R-NPSVM+ can produce a better performance, especially when the samples are corrupted by noise and outliers.

Automated summary

In this paper, we propose a robust nonparallel support vector machine model (R-NPSVM+) under the privileged information learning setting. R-NPSVM+ integrates privileged information into NPSVM to improve classification accuracy and uses robust loss functions to enhance model robustness. Experimental results demonstrate that R-NPSVM+ outperforms other classical algorithms, especially when samples are corrupted by noise and outliers.