Knowledge Source Selection by Estimating Distance between Datasets

Abstract

Most traditional machine learning methods make an assumption that the distribution of the training dataset is the same as the applied domain. Transfer learning omits this assumption and is able to transfer knowledge between different domains. It is a promising method to make machine learning technology become more practical. However, negative transfer can hurt the performance of the model, therefore, it should be avoided. In this paper, we focus on how to select a good knowledge source when there are multiple labelled datasets available. A method to estimate the divergence between two labelled datasets is given. In addition, we also provide a method to decide the mappings between features in different datasets. The experimental results show that the divergence estimated by our method is highly related to the performance of the model.