Probability Quantization Model for Sample-to-Sample Stochastic Sampling

Computer-driven sampling methodology has been widely used in various application scenarios, theoretical models and data preprocessing. However, these powerful models are either based on explicit probability models or adopt data-level generation rules, which are difficult to be applied to the realistic environment that the prior distribution knowledge is missing. Inspired by the density estimation methods and quantization techniques, a method based on the quantization modeling of data density named probability quantization model (PQM) is proposed in this paper. The method quantifies the complex density estimation model through the discrete probability blocks so as to approximate the cumulative distribution and finally achieve backtracking sampling based on the roulette wheel selection and local linear regression, which avoids the indispensable prior distribution knowledge and troublesome modeling constraints in the existing stochastic sampling methods. Meanwhile, our method based on the data probability modeling is hard to be influenced by the sample features, thus avoiding over-fitting and having better data diversity than over-sampling methods. Experimental results show the advantages of the proposed method to various computer-driven sampling methods in terms of accuracy, generalization performance and data diversity.