Incremental Discriminant Learning For Heterogeneous Domain Adaptation

In machine learning,a robust classifier often requires the use of large amounts of data to train.But in practice,collecting and labeling large amounts of data is expensive and time-consuming.Domain adaptation,as a kind of transfer learning method,enable the computer to transfer the classification model learned in one domain(source domain)to another domain(target domain),then completing the classification of the target data.In this method,the data in source domain is already labeled,the data in target domain is unlabeled or only small amount is labeled.Heterogeneous domain adaptation aims at learning a domain adaptation classifier when two domains have different feature representations.This paper proposes a heterogeneous domain adaptation method using incremental learning.Using two different projection matrices,through eigen calculation to map the features of the source domain and target domain into a common subspace with discrimination.In this subspace,the intra-class samples are closely-related to each other,the interclass samples are well-separated from each other,and the data distribution mismatch between the source and target domains is reduced.In order to be more applicable to scenarios with large amounts of data,this paper proposes a method of incremental learning.Data can come in batches or in the form of data streams,rather than all at once.Whenever new data arrives,our model will upgrade existing projection matrices to new projection matrices through eigen decomposition and eigen space merging algorithms,without using the entire data set to recalculate.Therefore,our incremental learning algorithm can significantly reduce time complexity.This paper conducts multiple sets of experiments on three different tasks: action recognition,object recognition and cross-linguistic text categorization.The experimental results prove the effectiveness of the proposed method in improving the accuracy of classification and reducing the calculation time.