Prospective domain adaptation for longitudinal data

Abstract

Longitudinal data includes the information of samples in various timepoints. When applying machine learning algorithms to this data, the use of up-to-date information will yield more accurate results. In this case, if the labels are derived from the up-to-date information, out-of-date samples for which data has not yet been collected are excluded from the prediction. To alleviate this problem, domain adaptation can be a method for the prediction of out-of-date samples in that the method transforms those features similar with the up-to-date samples and bridges to the use of labels. Especially, domain adversarial training methods with a gradient reversal layer derive feature representation where samples in different domains appear to be one set so as to make the origin of them indistinguishable. However, since existing methods focus on different data with heterogeneous features, by considering that homogeneous features are continuously collected in longitudinal data, those need to be improved for the out-of-date and up-to-date samples so that their features are exactly matched. Motivated by this, we propose a method of domain adaptation, namely prospective domain adaptation, where the transformation of the out-of-date features purposes to match the properties of the up-to-date features. Therefore, in the proposed method, the out-of-date features are adapted to the manifold and distribution of the up-to-date features so that two feature sets are implicitly and explicitly matched. The experimental results demonstrated that the proposed method derives well-matched feature representation and outperforms comparative methods.