Unsupervised Heterogeneous Transfer Learning for Partial Co-occurrence Data

Abstract

Heterogeneous transfer clustering contributes to improve the performance of target domain by using the co-occurrence data from different domains without any supervision. Existing works usually use a large of complete co-occurrence data to learn the projection functions mapping heterogeneous data to a common latent feature subspace. Given that in the real-world problems, complete and abundant co-occurrence data in the form of homogeneous transfer learning between the soured domain and target domain are hard to achieve, a heterogeneous transfer clustering method for partial co-occurrence data (HTCPC) is proposed here, to perform unsupervised learning to map the partial information obtained from the source domain onto objects in the target domain. Furthermore, to maximize the useful information to improve the clustering performance in target domain, the proposed HTCPC uses the deep matrix decomposition framework to maintain the multi-layer hidden feature representation and retain the complexity of the data hierarchy by adding the approximate orthogonal constraints, which can effectively strengthen the independence and minimal redundancy. From a series of experiments conducted on four datasets [Berkeley Drosophila Genome Project (BDGP), Devanagari Handwritten Character (DHC), Columbia University Image Library (COIL), and Notting-Hill (NH)], the results show that HTCPC outperforms the peers in the following aspects: our method constructs the hierarchical structure in the multi-layer latent representations and the proposed algorithm can reduce the redundancy and extract more useful knowledge for target domain.