Domain Adaptation on Time Series – DATeS

Machine learning models generally assume that training and test data are identically distributed. However, this assumption often breaks down in real-world applications due to shifting data distributions over time and across contexts. Domain Adaptation **(DA)** tackles this challenge by allowing models to leverage labeled data from one domain to perform tasks in a different, often unlabeled, target domain.

Time Series **(TS)** consists of data points recorded over time and is increasingly prevalent with autonomous sensors and online activities. In the context of domain adaptation, time series presents several challenges, including class imbalance, temporal drift, correlations and dependencies, and hierarchical structures that manifest as multi-scale periodic patterns.

The DATeS project addresses three fundamental challenges of domain adaptation for time series. The first challenge is efficiently achieving both temporal and domain alignment, which we will tackle by proposing new formulations based on generalized optimal transport and conditional flow matching. The second challenge involves adapting to missing data and source-free domain adaptation, where only a learned model is available; we will tackle this in the lens of optimal transport approaches. The third challenge relates to unsupervised domain adaptation (with few or no target labels) and involves the general problem of selecting the best model, source domain(s), and hyperparameters without access to a validation set. We will address this challenge by learning metrics tailored for time series in domain adaptation.

This project aims to push the boundaries in the field of domain adaptation for time series, addressing critical challenges that are currently under-explored. While DATeS is motivated by the challenges posed by time series data, it will also inspire innovative developments that have broader applicability beyond the time series domain.