Transfer Learning methods for temporal data

In this work, we propose novel transfer learning methods for time series analysis. Motivated by applications in household electricity consumption disaggregation (NILM) and industrial monitoring, we investigate traditional tools and show the need for specific transfer learning methods for time series. After reviewing existing transfer learning and domain adaptation frameworks, we formulate the following problem: can we extract transferable features from time series? Namely, our goal is to propose methods to reduce the performance gap when the test data comes from a different sample from the training data. In a first part, we investigate time series pre-processing for neural networks with a view on household consumption. We propose a normalization ensembling method that allows for better robustness when learning on devices from one house and applying on new devices.In a second part, we develop a general framework for adversarial domain adaptation for regression. We provide learning bounds and an algorithm for both single and multi-source domain adaptation. While our focus is on applications for regression, our framework is generic and is applicable to classification. We conduct extensive experiments on both private and public datasets, showing better performance than previous domain adaptation methods.Finally, we investigate transfer learning methods for multivariate time series using covariance information. We represent multivariate time series by their autocovariance matrix and develop a domain adaptation framework using the specific geometry of those matrices. Experiments show the efficiency of the method on synthetic data and promising results on Human Activity Recognition datasets.