Causal Discovery on Irregular Time Series: Extending PCMCI+ Beyond Fixed Lags

Forum topic · 小凯 · 2026-07-22

Summary

Researchers extend PCMCI+, a state-of-the-art method for causal discovery in regularly sampled multivariate time series, to handle irregularly sampled event streams. Standard causal discovery methods rely on regular, discrete lag structures, which limits their applicability to real-world data such as sensor readings, medical records, and financial transactions that arrive at irregular intervals. The proposed approach replaces fixed-lag dependency modeling with aggregation of causal influence over predefined temporal windows, making it robust to irregular sampling. The authors evaluate the method on synthetic irregular event streams with known ground-truth causal structures across varying signal-to-noise ratios. Results show the method consistently recovers the underlying causal graph and substantially outperforms standard PCMCI+ on irregularly sampled data. The paper is available on arXiv (2607.18226) under categories cs.LG and stat.ME.

Paper Overview

Research area: Machine Learning
Authors: Martim Penim, Ricardo Ribeiro Pereira, Jacopo Bono, Hugo Ferreira, Mário A. T. Figueiredo, et al. (6 authors)
Published: 2026-07-20
arXiv: 2607.18226
Categories: cs.LG, stat.ME

Abstract

Causal discovery methods have shown strong performance in time series systems, but they typically rely on regular, discrete lag structures, which restricts their applicability to regularly sampled data. However, many real-world tasks involve irregularly sampled event streams, such as sensor data, medical data, and financial transactions.

In this work, the authors extend PCMCI+ — a state-of-the-art method for causal discovery on regularly sampled multivariate time series — to handle irregular time series. Instead of modeling causal dependencies through fixed lags, the proposed method aggregates causal influence over predefined temporal windows.

Evaluation

The method was evaluated on synthetic irregular event streams with known ground-truth causal structures, across varying signal-to-noise ratios. Results demonstrate that it:

Consistently recovers the underlying causal graph
Significantly outperforms standard PCMCI+ on irregularly sampled data

Condensed Original Abstract

> We extend PCMCI+ to handle irregular time series by aggregating causal influence over predefined temporal windows instead of fixed-lag dependencies, substantially outperforming standard PCMCI+ on irregularly sampled data.

---

*Auto-collected on 2026-07-22*