An Improved Virtual Sample Generation Technology based on Mega Trend Diffusion

Abstract

The problem of small sample size is a major and severe problem in the era of big data. At present, the common solution to the problem of small samples is to expand the number of samples by using virtual sample generation (VSG) technology. But there are still many problems such as rationality, accuracy and practical application. Based on this, this paper proposes an improved VSG technology. At first, a region expansion technique based on Euclidean distance of sub-region is proposed, which extends the input sample space and output sample space. Then, in the extended region, the virtual input samples are generated equidistantly. The virtual output samples are generated by the mapping model and the virtual input samples, and the virtual output samples are deleted by combining the output sample space. Next, a random weight neural network (RWNN) is used to generate virtual input and output samples based on its hidden layer. Finally, these true and virtual samples are mixed to produce the final modeling. The validity and rationality of the VSG technology are validated by benchmark data set.