Extended nearest neighbor chain induced instance-weights for SVMs

• A novel instance-weighted strategy is proposed for SVMs.
• The samples near decision plane are assigned higher weights via the proposed method.
• The sample near decision plane can be found by a chain named extended nearest neighbor chain.
• The proposed method makes SVM become more robust for noises.

Prior knowledge plays an important role in increasing the performance of support vector machine (SVM). One way to utilize prior knowledge is to assign different weight to each sample, which is called weighted support vector machine (WSVM). In this paper, a novel instance-weighted method is proposed for support vector machines. The weight is determined by the distances between a sequence of the nearest heterogeneous samples, named the extended nearest neighbor chain. The samples near the decision plane also locate near the heterogeneous samples. Thus, the samples near the decision plane can be found by the extended nearest neighbor chain before SVMs training and then assigned higher weights. Meanwhile the samples far away from the decision plane are assigned lower weights. This is consistent with the fact that the samples near the decision plane are more important for SVMs. Moreover, this strategy can also assign lower weights to the samples which locate in the wrong classes. The results of the experiments, performed on both artificial synthetic datasets and real-world benchmark datasets, demonstrate that the proposed instance-weighted method can further improve the performance of SVMs.