An Incremental Principal Component Analysis based on dynamic accumulation ratio

We have proposed an online feature extraction method called Chunk Incremental Principal Component Analysis (CIPCA) where a chunk of data is trained at a time to update an eigenspace model. This paper presents an extended version in which the threshold for accumulation ratio is adaptively determined so that the classification accuracy for validation data is always maximized. To define the validation set online, the prototypes are selected from given training samples by k-means clustering or nearest neighbor classifier. The experimental results show that the proposed CIPCA can update the threshold properly so as to maintain high classification accuracy.