Estimating Mutual Information between Time Series and Temporal Event Sequences Across Diverse Analysis Tasks
Researchers propose a nonparametric mutual information estimator that quantifies dependence between continuous time series and discrete temporal event sequences without requiring data transformation or ad hoc discretization. The method addresses limitations in existing approaches through latent event clustering and continuous-discrete duality modeling, offering robust applications across causality analysis, pattern discovery, and feature selection tasks.
This research tackles a significant gap in temporal data analysis by developing a principled framework for measuring dependence between fundamentally different data types. Existing methods suffer from sensitivity to quantization artifacts and event redundancy, producing unreliable results when analyzing mixed data structures common in real-world applications. The proposed nonparametric estimator bridges discrete and continuous mutual information through clever architectural design that directly handles the inherent duality without lossy preprocessing steps.
The advancement builds on decades of dependence measurement research, from Pearson correlation for homogeneous data to modern causal inference techniques. However, heterogeneous temporal data—combining price movements with transaction timestamps, or sensor readings with event logs—has lacked a unified analytical approach. This innovation specifically addresses that limitation by eliminating the need for practitioners to choose between data transformation strategies, each introducing their own biases.
The practical implications span multiple domains reliant on temporal analysis. In quantitative finance, practitioners analyzing cryptocurrency price movements alongside blockchain events gain a more reliable causality detection tool. For time series forecasting, improved feature selection between continuous and discrete covariates enhances model accuracy. The framework's robustness against repeated values and event clustering artifacts makes it particularly valuable for real-world datasets prone to recording redundancy and sparse events.
Looking ahead, adoption depends on integration into popular data science libraries and validation across diverse industry datasets. The open-source code release facilitates this process, but widespread impact requires demonstrating clear performance advantages over ensemble or transformation-based approaches in production environments.
- →Proposes first principled nonparametric mutual information estimator for continuous-discrete temporal data without data transformation
- →Introduces latent event clustering strategy to eliminate bias from redundant or co-occurring events
- →Demonstrates consistent improvements across four distinct temporal analysis tasks including causality detection and feature selection
- →Positions method as general-purpose dependence operator comparable to Pearson correlation for heterogeneous data types
- →Open-source implementation available, enabling integration into existing temporal data analysis workflows