StreamStory: Exploring Multivariate Time Series on Multiple Scales.

Area of Science:

• Data Visualization
• Time Series Analysis
• Machine Learning

Background:

• Large multivariate time series data present challenges in pattern identification due to complexity and scale.
• Periodic and recurrent behaviors, often driven by variable interactions, are key patterns in such datasets.
• Existing methods may require significant data analytics expertise.

Purpose of the Study:

• To develop an interactive approach for visualizing, exploring, and interpreting large multivariate time series.
• To enable users, including non-experts, to identify patterns and understand temporal dynamics.
• To create a tool that abstracts complex data while retaining domain-specific interpretability.

Main Methods:

• Summarizing data into conceptual states and modeling temporal dynamics as state transitions.
• Extending data representation to multiple spatial granularities for multi-scale pattern discovery.
• Developing an interactive web-based tool (StreamStory) integrating abstraction with statistical and machine learning techniques.

Main Results:

• StreamStory effectively visualizes large datasets, potentially billions of examples.
• The tool facilitates identifying main system states, mapping them to data-specific concepts.
• High-level periodic behavior and multi-scale phenomena were successfully identified and interpreted across real-world datasets.

Conclusions:

• StreamStory provides an accessible and powerful method for visual analytics of complex time series data.
• The approach enables discovery of both known and novel patterns across different temporal scales.
• The tool empowers users with limited data analytics background to interpret intricate datasets.