Transfer Learning for Data Stream Mining in Non-Stationary Environments

Abstract

The relationship between the input and output data changes over time refer to as concept drift, which is a major problem in online learning due to its impact on the predictive performance of data stream mining systems. Typically, learning models cannot be well trained until sufficient data representing the new concept have been collected. Transferring the knowledge learnt from different sources to accelerate the learning process of the new target concept and to improve the predictive performance is a feasible direction. This thesis concentrates on how to use transfer learning to improve the predictive performance in non-stationary environments (e.g., learning concept drifts). The main contributions of this thesis consist of: • The first approach is able to transfer the knowledge between multiple data streaming sources in non-stationary environments. This approach is called Multi-sourcE onLine TrAnsfer learning for Non-statIonary Environments (Melanie). • The first approach is able to transfer the knowledge between multiple data sources even when source and target concepts do not match in nonstationary environments. This approach is called Multi-source mApping with tRansfer LearnIng for Non-stationary Environments (MARLINE). • The first two approaches are able to transfer the knowledge from different labels and label dependencies between multi-sources in nonstationary environments. These two approaches are called Binary Relevance Multi-Label classi?cation in non-stAtionary enviRonments with muLti-SourcE traNsfer lEarning (BR-MARLENE) and Binary Relevance PairWise Multi-Label classi?cation in non-stAtionary enviRonments with muLti-SourcE traNsfer lEarning (BRPW-MARLENE). • We launch comprehensive evaluations on the proposed methods against different methods with different datasets to have a better understanding of what contents, when and how transfer learning can help learning models to improve the performance in non-stationary environments.