Integrating Local Outlier Factor with Radial Basis Function Interpolation for Small Sample Size Problems

The accuracy of data-driven models requires large number of training samples. However, training samples in many applications are usually difficult to collect. Thus, learning from small training sets is still an unprecedented challenge. In this paper, we proposed an effective virtual sample generation approach integrating local outlier factor with radial basis function interpolation for solving small samples size problems. To capture underlying distribution of data points, local outlier factor is first employed to indicate the degree of outlierness for data points. Based on the degree, interesting points are over-sampled via a deliberately designed roulette wheel. Then virtual samples are synthetized by the RBF interpolation on the lines between an interesting point and its corresponding nearest neighbors. The experimental results over two benchmark function datasets and a real-world dataset demonstrate that the proposed approach outperforms the existing state-of-the-art methods.