A Fuzzy Set Based Classification-to-Regression Extension Framework for Transfer Learning and its Application to Brain-Computer Interfaces

Transfer learning (TL) uses data or knowledge from source domains to facilitate the learning in a target domain. It has achieved promising performance in many applications such as brain-computer interfaces (BCIs). However, there are much fewer TL algorithms for regression than those for classification. This paper addresses this challenge, by proposing a fuzzy set based framework to estimate class-conditional probabilities in regression. Using this framework, class-conditional probability-based traditional or deep TL algorithms for classification can be extended to regression. We further propose a novel TL algorithm that outperforms naive extensions of classification methods. Experiments on three electroencephalogram (EEG) based BCI regression datasets showed that the extended algorithms consistently outperformed several classic and state-of-the-art regression TL approaches, demonstrating the effectiveness and easiness of the fuzzy extension framework. Beyond TL, the fuzzy extension framework may also be applied to other machine learning problems when class-conditional information is needed.