A feature-based information-theoretic approach for detecting interpretable, long-timescale pairwise interactions from time series

• DOI: 10.48550/arxiv.2404.05929
• Publication Date: 2024-04-08
• Authors: Aria Nguyen; Oscar McMullin; Joseph T. Lizier; Ben Fulcher
• Keywords: Feature (linguistics)

Quantifying relationships between components of a complex system is critical to understanding the rich network interactions that characterize behavior system. Traditional methods for detecting pairwise dependence time series, such as Pearson correlation, Granger causality, and mutual information, are computed directly in space measured time-series values. But systems which mediated by statistical properties series (`time-series features') over longer timescales, this approach can fail capture underlying from limited noisy data, be challenging interpret. Addressing these issues, here we introduce an information-theoretic method features provides interpretable insights into nature interactions. Our extracts candidate set sliding windows source assesses their role mediating relationship values target process. Across simulations three different generative processes, demonstrate our feature-based outperform traditional inference based on raw values, especially scenarios characterized short lengths, high noise levels, long interaction timescales. work introduces new tool inferring interpreting feature-mediated contributing broader landscape quantitative analysis research, with potential applications various domains including but not neuroscience, finance, climate science, engineering.