Addressing Big Data Time Series: Mining Trillions of Time Series Subsequences Under Dynamic Time Warping.

Area of Science:

• Data Mining
• Machine Learning
• Time Series Analysis

Background:

• Time series data mining algorithms are often bottlenecked by similarity search.
• Current algorithms struggle to scale beyond millions of time series objects.
• Billions of time series objects in industry and science remain unexplored.

Purpose of the Study:

• To develop novel methods for mining massive time series datasets.
• To enable efficient similarity search for large-scale time series data.
• To address the scalability limitations of existing time series data mining algorithms.

Main Methods:

• A combination of four novel ideas for time series data mining.
• Exact search under Dynamic Time Warping (DTW) on large datasets.
• Demonstration on the largest time series datasets ever attempted, up to one trillion data points.

Main Results:

• Achieved significantly faster exact search under DTW compared to Euclidean distance on massive datasets.
• Successfully mined time series data at unprecedented scales.
• Demonstrated efficient solutions for higher-level problems like motif discovery and clustering.
• Showcased efficient support for the uniform scaling distance measure.

Conclusions:

• Novel methods enable the first-ever mining of massive time series datasets.
• The approach overcomes scalability limitations, opening new avenues for research and industry.
• Implications for real-time data stream monitoring and resource-constrained devices.