Transfer learning for one class SVM adaptation to limited data distribution change

Abstract

Data based one class classification rules are widely used in system monitoring. Due to maintenance for example, we may come across a change of data distribution with respect to training data. While lacking of representative samples for the new data set, one can try to adapt the former learned detection rule to the new data set instead of retraining a new rule which implies to gather a significant amount of data. Based on the above, a multi-task learning detection rule approach is proposed to deal with the training of the updated system as some new data are available. The key feature of the new approach is the introduction of a parameter to control how much we rely on the former model. This parameter has to be set and changed as the amount of new data coming from the system increases. We define the new detection model as a classical one class SVM with a specific kernel matrix which depends on the parameter we introduced. A kernel adaptation method for C-one class SVM is developed in order to get the path solution along that parameter and a criteria is established to select a good value. Experiments conducted on toy data and real data set show that the proposed method could adapt to data change, and it gives a good transition from the old detection rule to the new one which is just obtained using the new data set only when the number of samples gathered from that new one is large enough.