Semisupervised domain adaptation for mixture model based classifiers

This paper introduces a method for mixture model-based classifier domain adaptation, wherein one has adequate labeled training data for one (source) domain, very scarce labeled data for another (target) domain, and where the discrepancy between the source and target domain class-conditional distributions is not too great. Starting from the source domain classifier parameters, the method maximizes the likelihood of target domain data, while constrained to agree as much as possible with the target domain label information. This is achieved via an expectation maximization (EM) algorithm, where the joint distribution of the latent variables in the E-Step is parametrically constrained, in order to ensure space-partitioning implications are gleaned from the labeled target domain samples. Experiments on publicly available Internet packet-flow traffic data from different temporal and spatial domains demonstrate significant gains in classification performance compared to 1. direct porting of the source domain classifier; 2. semisupervised learning using only the target domain data; and 3. extension of an existing unsupervised domain adaptation method.