A SVM-based model-transferring method for heterogeneous domain adaptation

In many real classification scenarios the distribution of test (target) domain is different from the training (source) domain. The distribution shift between the source and target domains may cause the source classifier not to gain the expected accuracy on the target data. Domain adaptation has been introduced to solve the accuracy-dropping problem caused by distribution shift phenomenon between domains. In this paper, we study model-transferring methods as a practical branch of adaptation methods, which adapt the source classifier to new domains without using the source samples. We introduce a new SVM-based model-transferring method, in which a max-margin classifier is trained on labeled target samples and is adapted using the offset of the source classifier. We call it Heterogeneous Max-Margin Classifier Adaptation Method, abbreviated as HMCA. The main strength of HMCA is its applicability for heterogeneous domains where the source and target domains may have different feature types. This property is important because the previously proposed model-transferring methods do not provide any solution for heterogeneous problems. We also introduce a new similarity metric that reliably measures adaptability between two domains according to HMCA structure. In the situation that we have access to several source classifiers, the metric can be used to select the most appropriate one for adaptation. We test HMCA on two different computer vision problems (pedestrian detection and image classification). The experimental results show the advantage in accuracy rate for our approach in comparison to several baselines. (C) 2016 Elsevier Ltd. All rights reserved.