Abstract: Analyzing time series data is crucial to a wide spectrum of applications, including economics, online marketplaces, and human healthcare. In particular, time series classification plays an indispensable role in segmenting different phases in stock markets, predicting customer behavior, and classifying worker actions and engagement levels. These aspects contribute significantly to the advancement of automated decision-making and system optimization in real-world applications. However, there is a large consensus that time series data often suffers from domain shifts between training and test sets, which dramatically degrades the classification performance. Despite the success of (reversible) instance normalization in handling the domain shifts for time series regression tasks, its performance in classification is unsatisfactory. In this paper, we propose $\textit{FIC-TSC}$, a training framework for time series classification that leverages Fisher information as the constraint. We theoretically and empirically show this is an efficient and effective solution to guide the model converges toward flatter minima, which enhances its generalizability to distribution shifts. We rigorously evaluate our method on 30 UEA multivariate and 85 UCR univariate datasets. Our empirical results demonstrate the superiority of the proposed method over 14 recent state-of-the-art methods.

Lay Summary:

Time series data—like patterns in brain signals or motion sensor readings—is often used to classify different types of behaviors or conditions. For example, in healthcare, it can help distinguish between healthy and abnormal heart activity. However, models trained for this kind of classification often fail when applied to new data collected under slightly different conditions—a common real-world problem known as domain shift. Our work introduces a new method, called FIC-TSC, that helps machine learning models remain accurate even when the data changes. It does this by applying a principle from statistics called Fisher information, which encourages the model to learn more stable and reliable decision boundaries. We tested our approach on over 100 benchmark datasets and found it consistently outperformed most other leading methods. This makes our technique a strong candidate for real-world applications where reliability across different environments is essential.