Cross-Domain Metric andMultiple Kernel Learning Based on Information Theory

摘要

Learning an appropriate distance metric plays a substantial role in thernsuccess of many learning machines. Conventional metric learning algorithmsrnhave limited utility when the training and test samples are drawnrnfrom related but different domains (i.e., source domain and target domain).rnIn this letter, we propose two novel metric learning algorithmsrnfor domain adaptation in an information-theoretic setting, allowing forrndiscriminating power transfer and standard learning machine propagationrnacross two domains. In the first one, a cross-domain Mahalanobisrndistance is learned by combining three goals: reducing the distributionrndifference between different domains, preserving the geometry of targetrndomain data, and aligning the geometry of source domain data withrnlabel information. Furthermore, we devote our efforts to solving complexrndomain adaptation problems and go beyond linear cross-domain metricrnlearning by extending the first method to a multiple kernel learningrnframework. Aconvex combination of multiple kernels and a linear transformationrnare adaptively learned in a single optimization, which greatlyrnbenefits the exploration of prior knowledge and the description of datarncharacteristics. Comprehensive experiments in three real-world applicationsrn(face recognition, text classification, and object categorization) verifyrnthat the proposed methods outperform state-of-the-art metric learningrnand domain adaptation methods.