A New Multi-Layer Classification Method Based on Logistic Regression
Kai Kang, Fengqiang Gao, Junguo Feng · 2018
To improve the effect of logistic regression in multiobjective classification and explore its greatest potential, a set of training and classification algorithms is constructed, by using the high accuracy of two-class classification. Multi-layer predictions are made under the premise of ensuring clear structure of the model. The method of outlier detection is introduced to choose a proper number of two-class classifiers for categories that are prone to be confused. Then further predictions are made with these two-class classifiers. The evaluation on MNIST dataset show that this method can effectively improve the classification accuracy of multi-class datasets with limited increase of running time.