A constraint optimization approach to causal discovery from subsampled time series data

Abstract

We consider causal structure estimation from time series data in which measurements are obtained at a coarser timescale than the causal timescale of the underlying system. Previous work has shown that such subsampling can lead to significant errors about the system’s causal structure if not properly taken into account. In this paper, we first consider the search for system timescale causal structures that correspond to a given measurement timescale structure. We provide a constraint satisfaction procedure whose computational performance is several orders of magnitude better than previous approaches. We then consider finite-sample data as input, and propose the first constraint optimization approach for recovering system timescale causal structure. This algorithm optimally recovers from possible conflicts due to statistical errors. We then apply the method to real-world data, investigate the robustness and scalability of our method, consider further approaches to reduce underdetermination in the output, and perform an extensive comparison between different solvers on this inference problem. Overall, these advances build towards a full understanding of non-parametric estimation of system timescale causal structures from sub-sampled time series data.

Survey and Evaluation of Causal Discovery Methods for Time Series

“…For example, Gong et al (2015) considers the problem of subsampling which amounts to recover relations between time instants that were not observed as their difference is smaller than the sampling rate of the time series. Hyttinen et al (2017) further studies the subsampling in the context of time series with hidden variables. Gong et al (2017) studies methods to infer causal relations on time series which correspond to aggregate (local averages or sums of observations) of other time series.…”

Discovering causal graphs with cycles and latent confounders: An exact branch-and-bound approach

“…(2017) study the use of different priors. Also BIC approximations can be utilized (Hyttinen et al, 2017a). The approach of Claassen and Heskes (2012) obtains probabilities for d-separation relations by Bayesian model averaging over graphs.…”

“…In this work, we take on the challenge of improving the scalability of practical exact algorithms for the general search space of causal graph allowing for latent confounding variables and cycles. Recently, there has been noticeable interest in developing algorithmic solutions to this general problem setting and its variants (Triantafillou et al, 2010;Triantafilou et al, 2010;Hyttinen et al, 2013Hyttinen et al, , 2014Magliacane et al, 2016;Borboudakis and Tsamardinos, 2016;Zhalama et al, 2017;Hyttinen et al, 2017a). The first exact approach to the problem we focus on here was proposed in (Hyttinen et al, 2014), based on declaratively encoding the underlying optimization task as answer set programming (ASP) and applying an ASP solver to obtain provably optimal solutions to the problem.…”

Causal discovery algorithms: A practical guide

“…Time series data (e.g., a month‐by‐month series of measurements of an economy) provide additional constraints for causal inference since we have timing information. However, they also present a number of distinctive challenges and so can require quite different causal search algorithms (e.g., Entner & Hoyer, ; Hyttinen, Plis, Jarvisalo, Eberhardt, & Danks, ; Moneta, Chlaß, Entner, & Hoyer, ). We focus here exclusively on the static case of measurements of different individuals, as most data collected or analyzed by philosophers are of this form.…”