Mining Multivariate Correlations in Time Series Data

What is it about?

This work introduces novel class of multivariate relationship patterns called multipoles that could be very useful in correlation analysis of a time series dataset. Informally a multipole refers to a subset of sensors (that produce a time series) that have strong linear dependence (the paper proposes new measures to capture the strength of linear dependence) such that elimination of any one of the sensor from the set would lead to significant weakening of the linear dependence. The paper formally conceptualizes the notion of multipole as well as propose efficient algorithms to find such patterns from a large-scale time series dataset. High impactful applications of this work has been demonstrated in multiple domains including climate science and neuroscience.

Featured Image

Why is it important?

The tools proposed in this work could be highly relevant for exploratory analysis such as correlation analysis in a variety of domains where time series data is prevalent. Such domains could include financial stock market data, internet traffic data, flow network analysis, and also climate science and neuroscience. Specifically in climate science, this work could be highly relevant in mining novel multivariate teleconnections which are otherwise very difficult to explore using standard techniques used in climate science literature. One such teleconnection that was discovered by our work has already been scientifically analyzed and published into a top-tier climate science journal.

Perspectives