Abstract:In this paper, we introduce a novel class of graphical models for representing time lag specific causal relationships and independencies of multivariate time series with unobserved confounders. We completely characterize these graphs and show that they constitute proper subsets of the currently employed model classes. As we show, from the novel graphs one can thus draw stronger causal inferences -- without additional assumptions. We further introduce a graphical representation of Markov equivalence classes of the novel graphs. This graphical representation contains more causal knowledge than what current state-of-the-art causal discovery algorithms learn.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is how to use a new class of causal ancestral graphs (time - series DMAGs, i.e., ts - DMAGs) to represent causal relationships and independencies in multivariate time - series in the presence of latent variables. Specifically, the paper focuses on: 1. **Defining a new graphical model**: A new graphical model - time - series DMAGs (ts - DMAGs) is introduced to represent causal relationships and independencies in time - series data with latent variables. These graphs are obtained by projection from infinite time - series DAGs (ts - DAGs). 2. **Characterizing time - series DMAGs**: Time - series DMAGs are fully characterized and proven to be a strict subset of the currently used model classes. This means that stronger causal inferences can be drawn from time - series DMAGs without additional assumptions. 3. **Introducing time - series DPAGs**: Time - series DPAGs (ts - DPAGs) are proposed as a graphical representation of the Markov equivalence class of time - series DMAGs. These graphs contain more causal knowledge than the graphs learned by current causal discovery algorithms. ### Specific problems and solutions - **Problem background**: - **Limitations of existing methods**: Existing causal graph models (such as DAGs and MAGs) have limitations when dealing with time - series data, especially in the presence of latent variables. For example, the Granger causality framework requires the absence of latent variables and cannot handle contemporaneous causal influences. - **Characteristics of time - series data**: Time - series data has temporal order and potential latent variables, which makes it difficult for traditional causal graph models to be directly applied. - **Solutions**: - **Definition of time - series DMAGs**: Time - series DMAGs are obtained by projection from time - series DAGs, which describe the causal structure of time - series data. Time - series DMAGs can handle latent variables and maintain temporal order. - **Properties of time - series DMAGs**: The paper proves that time - series DMAGs are a new class of graphical models and a strict subset of existing model classes, so stronger causal inferences can be drawn from these graphs. - **Introduction of time - series DPAGs**: Time - series DPAGs are a graphical representation of the Markov equivalence class of time - series DMAGs, and these graphs contain more causal information than the graphs learned by existing causal discovery algorithms. ### Main contributions of the paper - **Theoretical contributions**: Time - series DMAGs are fully characterized and proven to be a strict subset of existing model classes, thus providing a more powerful tool for causal inference. - **Methodological contributions**: The concept of time - series DPAGs is proposed and an algorithm is provided to learn these graphs from data, thus improving the accuracy of causal discovery in practical applications. In summary, by introducing time - series DMAGs and time - series DPAGs, this paper provides new theoretical and methodological support for dealing with time - series data with latent variables, thus making important progress in the field of causal inference.

Characterization of causal ancestral graphs for time series with latent confounders

Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data

Inferring extended summary causal graphs from observational time series

Causal Graph Discovery from Self and Mutually Exciting Time Series

Mixed Graphical Models for Causal Analysis of Multi-modal Variables

Projecting infinite time series graphs to finite marginal graphs using number theory

Inferring Individual Level Causal Models from Graph-based Relational Time Series

Ancestor regression in structural vector autoregressive models

Causal Discovery under Latent Class Confounding

Normalized multivariate time series causality analysis and causal graph reconstruction

Identification of Latent Variables From Graphical Model Residuals

Causal Discovery of Linear Non-Gaussian Causal Models with Unobserved Confounding

Counterfactual Graph Learning for Anomaly Detection on Attributed Networks

Confidence in Causal Discovery with Linear Causal Models

Identifiability of total effects from abstractions of time series causal graphs

Entropy-based Discovery of Summary Causal Graphs in Time Series

Towards the Disappearing Truth: Fine-Grained Joint Causal Influences Learning with Hidden Variable-Driven Causal Hypergraphs in Time Series

GRACE-C: Generalized Rate Agnostic Causal Estimation via Constraints

Causal Discovery over High-Dimensional Structured Hypothesis Spaces with Causal Graph Partitioning

Multiple Causes: A Causal Graphical View

Constraint-based Causal Discovery for Non-Linear Structural Causal Models with Cycles and Latent Confounders