Online Transfer Learning for Mining Recurring Concept in Data Stream Classification

At the age of big data, data stream classification is being applied to many fields, like spam filtering, market predicting, and weather forecasting, et al, in which recurring concept is an important character. Aiming to reduce the influence of negative transfer and improve the lag of detection of concept drift, RC-OTL is proposed for mining recurring concepts in data stream based on online transfer learning strategy. When a concept drift is detected, RC-OTL selects one current base classifier to store, and then computes the domain similarities between the current training samples and the stored classifiers, in order to select the most appropriate source classifier to combine with a new classifier for learning the upcoming samples, which results in knowledge transfer from the source domain to the target domain. In addition, RC-OTL can select appropriate classifier to classify when the current classification accuracy is detected below a given threshold before concept drift detection. The preliminary theory analysis explains why RC-OTL can reduce negative transfer effectively, and the experiment results further illustrates that RC-OTL can efficiently promote the cumulate accuracy of data stream classification, and faster adapt to the samples of new concept after concept drift takes place.