Multi-Source Transfer Learning for Non-Stationary Environments

Abstract

In data stream mining, predictive models typically suffer drops in predictive\nperformance due to concept drift. As enough data representing the new concept\nmust be collected for the new concept to be well learnt, the predictive\nperformance of existing models usually takes some time to recover from concept\ndrift. To speed up recovery from concept drift and improve predictive\nperformance in data stream mining, this work proposes a novel approach called\nMulti-sourcE onLine TrAnsfer learning for Non-statIonary Environments\n(Melanie). Melanie is the first approach able to transfer knowledge between\nmultiple data streaming sources in non-stationary environments. It creates\nseveral sub-classifiers to learn different aspects from different source and\ntarget concepts over time. The sub-classifiers that match the current target\nconcept well are identified, and used to compose an ensemble for predicting\nexamples from the target concept. We evaluate Melanie on several synthetic data\nstreams containing different types of concept drift and on real world data\nstreams. The results indicate that Melanie can deal with a variety drifts and\nimprove predictive performance over existing data stream learning algorithms by\nmaking use of multiple sources.\n