UMN and Pitt Fix the Broken Tool Measuring How Events Shape Data Streams
A new nonparametric mutual information estimator bridges continuous time series and discrete event sequences, enabling causality analysis, seasonality detection, and feature selection without data transformation.
Takeaways
- New MI estimator: A nonparametric method directly measures dependence between continuous time series and discrete event sequences without binning or data transformation. Paper
- Two core innovations: Continuous-discrete duality handles repeated/quantized sensor values; latent event clustering merges redundant co-occurring event types to reduce bias.
- Fixes a known failure mode: Existing estimators (Ross, KSG) produce negative or wildly biased MI on quantized data; the new method is guaranteed non-negative and near-zero MSE on synthetic benchmarks.
- Validated on four tasks: Time-delayed causality, seasonality detection, discrete covariate selection for forecasting, and continuous feature selection for classification.
- Strong downstream results: With event clustering, wins 62–97/100 covariate ranking rounds vs. baselines across CatBoost, DeepAR, and Chronos-2 forecasters on Rossmann and M5 datasets.
- Open source: Full code at GitHub; core estimator in
normalize_clustered_mi.py.
Measuring how much a sensor reading depends on a system event sounds simple. In practice, it is one of the messiest open problems in time series analysis. You have two fundamentally different data types: a continuous stream of numbers (temperature, traffic volume, sales) and a discrete sequence of labeled events (a promotion started, a solar flare occurred, a holiday began). Standard tools like Pearson correlation only work when both sides are continuous. Mutual information estimators exist for mixed types, but they break down the moment your sensor rounds values to two decimal places or two events always co-occur. A new paper from the University of Minnesota and University of Pittsburgh, accepted at