A Feature Selection Method based on the Sparse Multi-Class SVM for Fingerprinting Localization

Feature selection in Fingerprint-based localization systems is of great importance, because of its capability to reduce the overhead in handling high-dimensional data while ensuring positioning accuracy. Several methods for such a task have been proposed, but they either do not consider the correlation between features, or propose an inefficient method to deal with the correlation. The study in this paper proposes a novel feature selection scheme based on the Sparse multi-class SVM (MSVM) technique which can address the feature selection problem via efficiently handling the correlation among them. The scheme first rules out several "unimportant" features via a simple criterion for scalability, and then selects a portion of the remaining features by controlling the sparsity of the optimization results of the Sparse MSVM. The method is applied to a realistic GSM-based fingerprinting localization system and the experimental results show that it outperforms several previous ones via reducing the mean localization error by about 20%.