Research for an Adaptive Classifier Based on Dynamic Graph Learning

Extreme Learning Machine (ELM) is a representative learning algorithm commonly used in data classification and prediction. In the previous literature on ELM, there are few works that pay attention to the relationship and geometric information of data, it may bring about the relatively low accuracy and poorer performance. Combining the classic ELM algorithm and the basic knowledge of dynamic graph learning, considering the geometric information between two data points to construct the graph matrix, an adaptive graph classifier on the basis of extreme learning machine is presented in our work. Besides, a matrix preserving the geometric information of the data is constructed from the original data and adaptively update during each training iteration. To do this, we use an alternative optimization strategy to update the graph matrix, so that the new classifier can adapt to the graph matrix and the graph matrix can update the classifier. The results on fifteen real data sets demonstrate that the proposed method outperforms in binary classification and multi-classification tasks.