Rethinking Chronological Causal Discovery with Signal Processing

summary

The gist

Causal discovery problems use a set of observations to deduce causality between variables in the real world, typically to answer questions about biological or physical systems In this paper…

In short

This work examines how causal discovery methods, like Granger Causality and Transfer Entropy, perform when dealing with time series data that might have mismatched sampling rates or window lengths compared to the true underlying physical process. The findings suggest that proper signal processing concepts can help improve robustness and guide the selection of crucial parameters for accurate causal inference.

Key concepts

Chronological Causal Discovery (CCD)
This task aims to find causal influences between different signals or time series by analyzing their chronological order. It is particularly important because it seeks to understand the generative process behind a dataset over time, often modeled using differential equations.
Signal Processing Concepts
Applying ideas from signal processing helps make causal discovery methods more robust against real-world data issues like subsampling (reduced sampling rate) and window length mismatches. This allows AI systems to better handle discrepancies between the observed data and the true continuous-time dynamics.
Nyquist Rate
This concept relates to signal processing and suggests a theoretical lower bound on the sampling rate needed to perfectly reconstruct a signal from its measurements. The paper explores whether a good notion of this causal Nyquist rate can be defined for causal inference.

Terminology used across episodes

This episode discusses

The paper

Rethinking Chronological Causal Discovery with Signal Processing · Read on arXiv

Kurt Butler, Damian Machlanski, Panagiotis Dimitrakopoulos, Sotirios A. Tsaftaris

School of Engineering, The University of Edinburgh

Causal discovery problems use a set of observations to deduce causality between variables in the real world, typically to answer questions about biological or physical systems. These observations are often recorded at regular time intervals, determined by a user or a machine, depending on the experiment design. There is generally no guarantee that the timing of these recordings matches the timing of the underlying biological or physical events. In this paper, we examine the sensitivity of causal discovery methods to this potential mismatch. We consider empirical and theoretical evidence to understand how causal discovery performance is impacted by changes of sampling rate and window length. We demonstrate that both classical and recent causal discovery methods exhibit sensitivity to these hyperparameters, and we discuss how ideas from signal processing may help us understand these phenomena.

DOI: 10.1109/IEEECONF67917.2025.11443432

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Rethinking Chronological Causal Discovery with Signal Processing".

Dev: Here's the gist of the paper. Causal discovery problems use a set of observations to deduce causality between variables in the real world, typically to answer questions about biological or physical systems In this paper, we examine the sensitivity of causal discovery methods to this potential mismatch We demonstrate that both classical and recent causal discovery methods exhibit sensitivity to these hyperparameters, and we discuss how ideas from signal processing may help us understand these phenomena Causal information is important because it allows us to reason about the world and the effects of how we interact with the world One such task is causal discovery, a detection-theoretic task in which one attempts to perform graph topology inference of a causal system In this work, our particular interest is in the chronological causal discovery (CCD) task, where the goal is to detect causal influences between signals or time series The gold standard for modelling how natural systems evolve in time, and by extension their causal relationships, is differential equations Many recent works in the causal machine learning (causal ML) literature have focused on discrete-time models, without questioning how faithful these models are to the underlying continuous-time dynamics If we consider the CCD task to be a problem of understanding the generative process behind a data set, then the mismatch between discrete and continuous becomes clearly important Review of Causal Discovery Methods The paper reviews several approaches to the chronological causal discovery task, categorized by their output graph structure 1) Granger Causality (GC) and Transfer Entropy (TE): GC is potentially the oldest approach to detecting causal influences TE formulates detection using information theory instead of predictive performance, measuring directed information flow 2) Constraint-based methods: These are based on conditional independence (CI) tests to recover a graph skeleton 3) Functional-based methods: These approaches assume stronger assumptions pertaining to the functional form of the data generating process, leading to identifiability guarantees 4) Gradient-based methods: NOTEARS casts the (acyclic) graph learning task into the continuous space, enabling the use of conventional optimizers 5) Cross Mapping: These methods assume that observed time series were generated by a latent deterministic dynamical systems Temporal Robustness and Hyperparameter Sensitivity One challenge of causal reasoning with statistical models is that there is an important distinction between the true data generative process (DGP) and our model of that process 1) A case study from neuroscience showed that for small values of Q, only a weak causal influence could be detected 2) Effects of sampling rate: subsampling in excess (↓ fS) makes the performance of causal methods deteriorate 3) Parameter selection: preferences for how to select Q are often method dependent, and extreme choices of Q may not yield good results Empirical Benchmarking Results The paper examines empirical evidence that proper selection of τ and Q matters for the CCD task 1) GC - Varying Q and k: based on GC show that causality cannot reliably be detected when Q < 50 TE was only able to detect causation when Q was matched to the propagation delay of the underlying causal signal 2) Benchmarking ML CCD: From both cases it is clear that varying hyperparameters Q and k has a large impact on the inferred causalities The right values for Q and k that lead to the right graph (dark red) also depend on used CCD algorithm Our observations point to challenges that CCD methods face when applied to real data The signal processing community has clear opportunities to contribute to the overcoming these challenges and developing better CCD methods There remains a question about whether a good notion of a causal Nyquist rate can be defined Acknowledgements We thank several individuals for useful discussions: Emily Stephen, Urbashi Mitra, Michael Camilleri, and Petar Djuric References

1: B.

Rosa: Scholkopf, F.

Dev: Locatello, S.

Rosa: Bauer, N.

Title and authors: Rosa: We just talked about how this paper, "Rethinking Chronological Causal Discovery with Signal Processing," investigates the sensitivity of causal discovery methods to timing mismatches in observations and the impact of sampling rate and window length.

Dev: Exactly, and they point out that we need to look at ideas from signal processing to build more robust ways to handle these discrepancies when we’re trying to find causality in time series.

Taro: It seems like a big deal because it suggests that just using standard methods might not be enough when the data collection isn't perfectly synchronized with the underlying process.

Rosa: They review several approaches, from things like Granger Causality and Transfer Entropy to constraint-based and functional-based methods, showing how each one reacts differently to these timing issues.

Dev: That’s important because it highlights a gap between modeling the data and truly understanding the underlying dynamics of a real system over time.

Taro: So, what they are suggesting is that we need to think more about signal processing concepts to bridge that gap in causal reasoning for time series.

The paper's summary: Rosa: The paper summarizes the findings by looking at specific examples, like a case study from neuroscience where only weak causal influences can be detected when the parameter Q is small.

Dev: They also point out that subsampling in excess, which is when you take too many samples or the sampling rate is too low, tends to make the performance of these causal methods deteriorate significantly.

Taro: One of their specific findings is that for Granger Causality, they found causality just couldn't be reliably detected when Q was less than fifty and Transfer Entropy only detected causation when Q matched the propagation delay of the signal.

Rosa: So, it’s not just a general sensitivity; it’s showing concrete numbers about how much performance drops based on those specific values for Q and k in different methods.

Dev: And they also showed that benchmarking causal machine learning for chronological discovery confirms that varying Q and k has a large impact on the inferred causalities, and even the right settings depend on the algorithm used.

Taro: That tells us that just picking a method isn't enough; you have to consider how well your chosen parameters align with the actual dynamics of what you are measuring.

The paper's improvements: Rosa: The paper suggests several ways to improve these methods, focusing on incorporating signal processing concepts to handle the mismatch between discrete and continuous data.

Dev: They suggest using concepts from signal processing to help us understand these phenomena better, which is a big idea because it moves beyond just tweaking parameters.

Taro: One improvement they propose is defining what they call a causal Nyquist rate, which would give us a proper way to define the lower bound on sampling rates needed for correct causal inference.

Rosa: That’s interesting because it means we could potentially use the spectral properties of the hypothesized causal mechanism to automatically select optimal sampling rates.

Dev: They also discuss how parameter selection methods, like choosing Q, can be improved by integrating information criteria such as Akaike or Bayesian criteria more deeply into the algorithm's decision-making process.

Taro: And they look at hybrid approaches, combining constraint-based methods with gradient-based techniques to get the best of both worlds—structural constraints and continuous optimization like DYNOTEARS.

Conclusion: Rosa: So, to wrap up, the paper on "Rethinking Chronological Causal Discovery with Signal Processing" really points out that when you apply these methods to real data, you run into challenges because of the mismatch between our model and the true data generating process.

Dev: They show that both classical and recent causal discovery methods are sensitive to things like sampling rate and window length, proving that we have some serious limitations when we don't account for those aspects.

Taro: The main implication is that the signal processing community has clear opportunities to contribute by developing better chronological causal discovery methods, especially by defining a good notion of a causal Nyquist rate.

Rosa: It suggests there’s a way forward where AI systems can perform more robust causal inference when data acquisition times don't perfectly match the underlying physical events.

Dev: We need to keep pushing on these ideas as we try to build systems that can reason about the world with more accuracy in dynamic situations, which sets us up for what comes next.

More episodes

← Home