Title: Robust and graph regularised non-negative matrix factorisation for heterogeneous co-transfer clustering

Authors: Yu Ma; Zhikui Chen; Xiru Qiu; Liang Zhao

Addresses: School of Software Engineering, Dalian University of Technology, Dalian 116620, China ' School of Software Engineering, Dalian University of Technology, Dalian 116620, China ' School of Software Engineering, Dalian University of Technology, Dalian 116620, China ' School of Software Engineering, Dalian University of Technology, Dalian 116620, China

Abstract: Transferring learning is proposed to tackle the problem where target instances are scarce to train an accurate model. Most existing transferring learning algorithms are designed for supervised learning and cannot obtain transferring results on multiple heterogeneous domains simultaneously. Moreover, the performance of transfer learning can be seriously degraded with the appearance of noises and corruptions. In this paper, a robust non-negative collective matrix factorisation model is proposed for heterogeneous co-transfer clustering which introduces the error matrices to capture the sparsely distributed noises. The heterogeneous clustering tasks are handled simultaneously and the graph regularisation is enforced on the collective matrix factorisation model to keep the intrinsic geometric structure of different domains. Experiment results on the real-world dataset show the proposed algorithm outperforms the baselines.

Keywords: transfer learning; non-negative matrix factorisation; NMF; error matrix; graph regularisation; clustering.

DOI: 10.1504/IJCSE.2019.10017878

International Journal of Computational Science and Engineering, 2019 Vol.18 No.1, pp.29 - 38

Received: 22 May 2017
Accepted: 12 Jul 2017

Published online: 14 Dec 2018 *

Full-text access for editors Access for subscribers Purchase this article Comment on this article