Abstract:

To solve the difficulties of high calculation quantity and low precision in constructing sparse Markov network  with a small set of samples, an iterative noise reduction (INR) algorithm based on the Gaussian noise model is  proposed. The algorithm firstly picks out the related variables through employing statistic test to regression residuals.  After that, a learning ability is gradually improved through the autoregressive update strategy similar as boosting  method. Finally, Akaike information criterion (AIC) is used to avoid overfit. In addition, the iterative update formula  is provided and the error rate controlling is realized. Furthermore, the computational complexity of the proposed  algorithm is analyzed. The experimental results show that INR can effectively construct the high dimensional sparse  network and has obvious advantages on learning precision and efficiency.