Scalable visual exploration of time series and anomalies with adaptive level of detail

Abstract

Time series datasets often contain numerous time series or segments that cannot be effectively displayed on a single screen, leading to information overload. This work addresses this challenge by integrating hierarchical navigation and semantic zooming techniques to enhance dataset exploration without overwhelming users. Our approach emphasizes scalability through hierarchical clustering, which allows efficient navigation of pattern groups. To inspect a specific pattern group, we employ a semantic zooming technique that combines three complementary visualizations: line charts, horizon graphs, and heatmaps. While our approach is generic, we also tailor it to support visual analysis of anomalies in time series, an important domain benefiting in particular from semantic zoom. We provide a visual anomaly score indicator that summarizes anomaly locations within clusters of time series. Additionally, a glyph visualizes each algorithm’s sensitivity for each cluster and time series. The practical value of our unique approach is validated through an ICE-T user study and several case studies, demonstrating its effectiveness for time series data exploration and analysis.