Transfer Learning with Copulas
David López-Paz and José Miguel Hernánez Lobato
In: NIPS 2011 Workshop on Copulas in Machine Learning, 16 Dec 2011, Granada, Spain.
Classic machine learning algorithms assume that both training and test data share the same underlying generative mechanisms. By contrast, transfer learning techniques relax this constraint by re-using knowledge obtained from other learning tasks in order to improve predictive performance. In this work, we propose a new transfer learning framework based on the assumption of a common copula model across different tasks. Copulas are statistical objects that fully describe the dependence structure of the data and link arbitrary marginal distributions into a complete multivariate model. We validate the effectiveness of our method in experiments with both synthetic and real-world data.