Transfer Learning with Part-Based Ensembles

Transfer learning is one of the most important directions in current machine learning research. In this paper, we propose a new learning framework called Multi-source part-based Transfer Learning (MspbTL), which is one kind of parameter transfer with multiple related source tasks. Dissimilar to many traditional works, we consider how to transfer information from one task to another in the form of integrating transferred information between parts. We regard all the complex tasks as a collection of several constituent parts respectively. It means that transfer learning between two complex tasks can be accomplished by subtransfer learning between their parts. Then, after completing the above information transfer between the source and target tasks, we integrate the models of all the parts in the target task into a whole. Experiments on some real data sets with support vector machines (SVMs) validate the effectiveness of our proposed learning frameworks.