Transfer learning for Bayesian networks

点击翻译按钮获取中文摘要

Transfer learning addresses the problem of how to learn a model for a domain (target domain) by utilizing an existing model from another different but related domain (source domain). The key issue is how to adapt the existing knowledge in source domain to fit target domain. Thus the representation of domain knowledge plays an important role. A Bayesian network (BN) is an explicit graphical representation of domain knowledge. It is natural to perform manipulations on a BN for source domain (source BN) to obtain a new BN for target domain (target BN). Therefore, BNs are suitable in the scenario of transfer learning. On the other hand, although BNs have been intensively studied for decades, learning BNs still remains quite challenging, especially when training data is scarce. The transfer learning approach can provide an effective way to overcome this issue in learning BNs. This proposed research will systematically study the transfer learning methods for BNs. We will first examine the suitability of the existing transfer learning approaches for the task of learning BNs, such as instance transfer, feature representation transfer, parameter transfer, etc. Then novel transfer learning methods that fit the task of learning BNs will be studied and developed. In addition, effective transfer learning algorithms for BNs, including both structure learning and parameter learning, will be studied. The evaluation of learning algorithms in this research is based on experiments on the commonly used benchmark data sets. The existing research in this field is in an early stage. The novelty and significance of this research are reflected in the following aspects: (1). BNs have not been well noticed in the scenario of transfer learning. The proposed research will demonstrate BNs as an effective transfer learning approach. (2). Structure learning is still a challenging issue in learning BNs. This research will provide a new approach to alleviate the structure learning issue, and thus develop more effective and efficient BN learning algorithms. (3). This research is expected to produce effective and efficient novel methods and algorithms for real-world data mining applications.