Marginalized Stacked Denoising Autoencoder With Adaptive Noise Probability for Cross Domain Classification

Cross-domain classification is a challenging problem, in which, how to learn domain invariant features is critical. Recently, significant improvements to this problem have emerged with the wide application of deep learning models, which have been proposed to learn higher level and robust feature representation. Marginalized stacked denoising autoencoder model (mSDA) has proved to be effective to address this problem. However, the performance of mSDA is sensitive to the noise probability. In previous works, the noise probability is usually set as a constant value by cross-validation in the source domain. There is few work focus on the relationship between the noise probability and cross-domain task. In this paper, we try to compute the value of noise probability adaptively. Thus, an approach called Marginalized Stacked Denoising Autoencoders with Adaptive noise Probability (mSDA-AP) is proposed. Firstly, we extract an informative feature space by an improved index, weighted log-likehood ratio (IWLLR), then aggregate these informative features by weighting. Secondly, we compute the value of noise probability adaptively according to the distance between source domain and target domain, and then with the adaptive noise probability, we disturb the input data to learn a stronger feature space with mSDA. Finally, experimental results show the effectiveness of our proposed approach.