Cross Domain Transfer Learning In Multimedia Data Mining

In the applications of data mining and machine learning, the supervised learning models are very useful to exploring the knowledge underlying in the data. However, there is a need of sufficient labeled data to train these models. In real world applications, we always lack the supervised information. It costs much labor and time to collect these labeled data. To deal with these problem, transfer learning methods are proposed. Transfer learning methods are able to utilize the information and knowledge from other relevant datasets, which is called as auxiliary domain, to help the learning process in the dataset of interest, which is denoted as target domain. In this paper, based on the deep investigation of the existing literature, we propose several novel models and algorithms to tackle with the unsolved problem in transfer learning. The key contributions of this paper are highlighted as follows:1. We develop a new unified framework called DTLM (Discriminative Transfer Learning on Manifold) to tackle with the transfer learning. We observe that in the existing literature of collective matrix factorization based transfer learning, the learned latent factors still suffer from the domain divergence and thus are usually not discriminative for an appropriate assignment of category labels, resulting in a series of issues that are either not addressed well or ignored completely. To address these issues, we have developed a novel transfer learning framework called DTLM. Specifically, based on a cross-domain matrix tri-factorization framework, we simultaneously incorporate a discriminative regression model and optimize the latent factors by minimizing the MMD distance between the different domains.2. By the intuition of transferring knowledge based on data instances, we propose a new transfer learning model called COSUF (cross domain shared subspace learning for unsupervised transfer classification). In the subspace, the innate characteristics of the concept in the dataset are explored and the original feature divergence is alleviated by minimizing the distance between different domains.3. We propose a new framework to discriminatively select the multi-view features for transfer learning. In the framework, we sparsely select the representative features shared by the semantics of the same class-label across different domains. To mitigate the gap between the domains, we adopt the Maximum Mean Discrepancy (MMD) criterion as a regularizer on the feature mapping coefficient matrix. Moreover, to enhance the predictive power in the target domain, we not only encode the local manifold structures of the data distribution into the predicted label matrix through a geometric graph regularization, but also incorporate the cross-domain global discriminative information to avoid overfitting and to make the predicted results more accurate.