Inductive transfer through neural network error and dataset regrouping

A new inductive transfer-learning algorithm called NEDRT is presented in this paper in order to improve the classification accuracy of a domain task by using the knowledge learned from labeled data generated from a different domain. NEDRT introduces a novel error function for a constructed neural network by summing a weighted squared difference between the real output and the neural network output for each instance of label training data from the source domain and the target domain. Each weight could be regarded as an instance's contribution degree to transfer, The source data set is partitioned into different sunsets to minimize the imbalance between the target data and source data, and each subset is combined with the target data to form a new training data set. These newly obtained training data sets are used to construct classifiers for the target task. Experimental results of knowledge transfer on UCI data sets and text data sets show that NEDRT performs well.