A Particle Swarm Optimization Based Joint Geometrical and Statistical Alignment Approach with Laplacian Regularization

Abstract

Transfer Learning or Domain Adaptation is an emerging sub-field of Machine learning in which the source domain carrying ample amount of labeled data is employed to classify a diverse but inter-related target domain data. However, in primitive machine learning algorithms, there is a pre-assumption that training and test data belong to the same set of distributions. But in a real-world scenario, the source (or training data), as well as the target domain (or test data), has a diverse distribution of data. Also, the source domain is equipped with the labeled data information while the target domain has a scarcity of labeled data or has completely unlabelled data information. Although the existing transfer learning algorithms like Joint Geometrical and Statistical Alignment (JGSA) takes into account several objectives to reduce the geometrical shift and distribution shift simultaneously, still they lack in fulfilling objectives like Laplacian regularization and degenerate feature elimination. So, we proposed a novel approach called Particle Swarm Optimization based Joint Geometrical and Statistical Alignment approach with Laplacian regularization (PSO-JGSAL) which incorporates new objective with JGSA and also utilizes the PSO approach for selection of an optimum set of features for classification purposes. Rigorous experiments have been done on Office-Caltech datasets (with SURF Features and Decaf Features) and our proposed PSO-JGSAL shows promising results as compared to already existing primitive and domain adaptation methods.