Non-Negative Collective Matrix Factorization Algorithm For Heterogeneous Co-Transfer Clustering

Transfer learning is proposed to solve the problem where the target data is scarce for learning an accurate model in traditional machine learning.The main idea of transfer learning is to leverage the amount of auxiliary data in the source domain to help the learning of target tasks.However,most existing transfer learning methods are designed for supervised learning and usually assume the source domain have the same feature space with the target domain.It cannot handle the knowledge transferring problem in multiple domains simultaneously.Thus,a non-negative collective matrix factorization method for heterogeneous transfer learning is proposed in this paper for solving the unsupervised transfer learning problem among heterogeneous domains.In order to handle the problem of data noises in the source and target domains,error matrix with ₁-norm constraint is firstly introduced to the traditional non-negative matrix factorization model to capture the sparsely distributed noises and corruptions among the original data.And this will make the learned common feature space more accurate.Moreover,in order to keep the intrinsic geometric structure of data in each domain,the manifold constraint is enforced on the robust non-negative matrix factorization model to control the smoothness of the process of the collective matrix factorization.And this is used to avoid the destruction of the intrinsic geometric structure of data in each domain due to the knowledge transferring.Base on the above improvements,the graph regularized robust non-negative matrix factorization model is extended to the format of collective matrix factorization,and the co-occurrence relationship among data is used to map the data from several heterogeneous domains to a common feature space,which serves as the bridge of knowledge transferring among heterogeneous domains.Besides,an iterative updating strategy is proposed in this paper to solve the above non-negative collective matrix factorization problem,and the detailed derivation processes about each updating rules are presented in this paper.In order to validate the effectiveness of the proposed heterogeneous co-transfer learning algorithm,elaborate experimental evaluation and comparison over two real-world data set are conducted as to some classical single-domain clustering algorithms and representative heterogeneous transfer learning method.The experimental results show that the proposed algorithm for heterogeneous co-transfer clustering achieves better performance and more robustness compared with other methods,which could effectively solve the unsupervised transfer learning problem among heterogeneous domains.