Online Adaptive Kernel Learning with Random Features for Large-scale Nonlinear Classification

Highlights：

• A novel random feature map is provided to improve the flexibility of kernel function, which can make the training samples linearly or approximately linear separable in high dimensional feature space.

• Random features based online adaptive kernel learning is proposed to deal with large-scale nonlinear classification, which guarantees the learning model can better adapt to the change of data distribution shape when data is coming one by one.

• The experiment results are conducted on twelve data sets and the results show that the proposed algorithm outperforms the state-of-the-art online methods on most data sets. Besides, the test accuracy of RF-OAK is comparable with that of offline deep learning algorithm on most data sets.