An Empirical Study of Adversarial Domain Adaptation on Time Series Data

Domain-adversarial learning allows a machine learning model to be trained with supplementary data from a different domain. This enables applications in various time series domains. Although several domain-adversarial models have been proposed in the past, there is a lack of empirical results with different types of time series. This paper provides an empirical analysis with multiple models, datasets and evaluation objectives. Two models known from literature are evaluated in combination with four public datasets: An RNN-based model (VRADA) is contrasted with a newer CNN-based one (CoDATS). The datasets include indoor climate, gas sensors, human activity and physiological data. Our experiments explicitly consider different dataset sizes, similarities between domains and the scenario of multisource training. It is found that CoDATS is very suitable for univariate datasets and performs well even with small datasets. Multivariate datasets can only benefit from the adversarial domain adaptation if the number of data points is large enough. VRADA was found to outperform CoDATS in modeling multivariate datasets. The multisource training available in CoDATS appears promising. A correlation is shown between the similarity of domains and prediction performance.