A particle swarm optimization-based feature selection for unsupervised transfer learning

Transfer learning (TL) method has captured an attractive presence because it facilitates the learning ability in the target domain by acquiring knowledge from well-established source domains. To gain strong knowledge from the source domain, it is important to narrow down the distribution difference between the source and the target domains. For this purpose, it is necessary to consider the objectives such as preserving the discriminative information, preserving the original similarity of the source and the target domain data, maximizing the variance of the target domain, and preserving marginal and conditional distribution at the same time. Furthermore, some existing TL methods use only original feature data, so there is a threat of degenerated feature transformation. To overcome all these limitations, in this paper, a novel feature selection-based transfer learning approach using particle swarm optimization (PSO) for unsupervised transfer learning (FSUTL-PSO) is implemented. In FSUTL-PSO, we incorporate all such objectives into one fitness function and select common good features from the source and target domains based on the fitness function for eliminating the threat of degenerated features. Extensive experiments have been done on all possible tasks of Office+Caltech and PIE Face datasets and our proposed method FSUTL-PSO has shown significant improvement over the existing transfer or non-transfer learning methods.