Transfer Learning For Bayesian Network Parameter

Parameter learning is crucial for the performance of Bayesian network.Parameter learning is to learn the CPTs of Bayesian network from train sample under the condition that the structure of Bayesian network is given.Maximum likelihood estimation is a classical and effective method for Bayesian Network parameter learning,but the maximum likelihood estimator is not consistent when learning from a small size of train sample.When training on small size of train sample,parameter learning could be conducted via constrained maximum likelihood estimation based on prior knowledge,which is hard to get in practice.Transfer learning is a new learning method which applies knowledge from different but related domains to the target domain.The goal of transfer learning is to improve the learning process in target domain via knowledge transfer when train sample from target domain is scarce.In this article,we proposed a new method,TL-WMLE,to solve the problem of learning parameters from small size of train sample and little prior knowledge via transfer learning by re-weighting the training sample from the source domain.The novel method compute the weight using an auxiliary classifier combining the SMOTE-N method and compute the weight of source sample depending on the predicted probability of being from the source domain by the auxiliary classifier.The re-weighted source train sample is then used together with the target train sample to build a likelihood function on the target domain for maximum likelihood estimation.Experiments on artificial data,standard data and text sentiment classification data show that TL-WMLE provides an effective and efficient way for learning Bayesian network parameters on small size of train sample and little prior knowledge.