Classification of time-series data with payload information
Elad Yom‐Tov · 2025
Time series data with rich payloads-such as text or images-consist of sequences where each sample contains high-dimensional content, and where the temporal aspect may be of relevance to the task at hand. Despite their importance, especially in domains such as healthcare, there are currently no effective methods for classifying such data, especially when samples are limited and time-series are sparse. We propose a novel classification approach that first embeds high-dimensional payloads into lower-dimensional representations, then applies sparse Dynamic Time Warping (DTW) in combination with support vector machines (SVMs) for classification. We evaluate this method on six Reddit datasets involving medically relevant transitions (e.g., from depression to suicidal ideation). Our results show a significant improvement in classification accuracy over an average pooling baseline. We also analyze the algorithm's performance in relation to dataset characteristics and model behavior, and we introduce an approach for interpreting the learned model. Overall, our method achieves superior classification performance in 5 of the 6 datasets and offers model interpretability, making it well-suited for medical and other high-stakes applications involving complex sequential data.