A Non-parametric Conformity-Based Test for Transfer Decisions

This paper proposes a new non-parametric test to decide whether to transfer from source data to target data in order to improve the performance of predictive models on target domains. The test is based on the conformity framework. It statistically tests whether the target data and source data have been generated from the target distribution under the exchangeability assumption. The source data is transferred if and only if the test is positive. The experiments show that the test is better at deciding on instance transfer than existing methods.