Instance transfer aims at improving prediction models for a target domain by transferring data from related source domains. The effectiveness of instance transfer depends on the relevance of source data to the target domain. When the relevance of source data is limited, the only option is to select a subset of source data of which the relevance is acceptable. In this paper, we introduce three algorithms that perform source-subset selection prior to model training. The algorithms employ a conformity-based test that estimates the source-subset relevance based on individual instances or on subsets as a whole. Experiments conducted on four real-world data sets demonstrated the effectiveness of the proposed algorithms. Especially, it was shown that pre-training subset-selection based on set relevance is capable of outperforming the existing instance-transfer techniques. (C) 2017 Elsevier B.V. All rights reserved.
|Number of pages||11|
|Publication status||Published - 4 Oct 2017|
- Conformal test
- Instance-transfer learning
- Source-subset selection