[Paper Review] Algebraic causality: Bayes nets and beyond

Eva Riccomagno, Jim Q. Smith|ArXiv.org|Sep 21, 2007

Bayesian Modeling and Causal Inference18 references3 citations

TL;DR

This paper proposes a general algebraic framework for causal modeling that extends Bayesian networks by formalizing causality through partial orders and polynomial maps, enabling the analysis of causal effects in non-graphical models. The key contribution is a unified algebraic approach that subsumes Bayesian networks, context-specific networks, and chain event graphs, allowing identifiability and feasibility to be studied via computational commutative algebra and semi-algebraic constraints.

ABSTRACT

The relationship between algebraic geometry and the inferential framework of the Bayesian Networks with hidden variables has now been fruitfully explored and exploited by a number of authors. More recently the algebraic formulation of Causal Bayesian Networks has also been investigated in this context. After reviewing these newer relationships, we proceed to demonstrate that many of the ideas embodied in the concept of a ``causal model'' can be more generally expressed directly in terms of a partial order and a family of polynomial maps. The more conventional graphical constructions, when available, remain a powerful tool.

Motivation & Objective

• To generalize the concept of causal models beyond graphical structures like Bayesian networks by embedding causality in algebraic and order-theoretic foundations.
• To address identifiability and feasibility of causal effects in complex models where traditional graphical representations are insufficient or cumbersome.
• To unify diverse model classes—such as context-specific Bayesian networks, Bayes Linear Constraint models, and Chain Event Graphs—under a single algebraic formalism.
• To demonstrate that causal inference can be initiated not from a graph, but from a finite set of unfolding events and a hypothesized causal order, enabling more flexible modeling.
• To show that algebraic techniques, including toric ideals and semi-algebraic constraints, can be systematically applied to analyze causal hypotheses and manipulated distributions.

Proposed method

• Represent causal models as a partial order on a finite set of events and a family of polynomial maps from one semi-algebraic space to another.
• Formalize causal interventions as projections of the joint probability mass function under the uncontrolled system, analogous to do-calculus in CBNs.
• Use parametrizations of conditional independence relations as polynomials, linking them to algebraic geometry concepts such as toric varieties and ideals.
• Model structural constraints (e.g., independence, conditional independence) as semi-algebraic equations and inequalities in the parameter space.
• Apply computational commutative algebra tools—such as Gröbner bases and elimination theory—to solve identifiability and feasibility problems.
• Illustrate the framework on a causal model of violent behavior influenced by movie exposure and testosterone levels, using three experimental data sources to estimate total and direct effects.

Experimental results

Research questions

• RQ1 How can causal models be formalized independently of graphical representations using algebraic and order-theoretic structures?
• RQ2 In what ways can the identifiability of causal effects be analyzed using polynomial constraints and semi-algebraic geometry?
• RQ3 Can the framework of causal Bayesian networks be generalized to include non-graphical models such as context-specific networks and Chain Event Graphs?
• RQ4 What are the algebraic properties that characterize the joint distribution under intervention, and how can they be derived from the underlying causal order?
• RQ5 How do different experimental data configurations (e.g., marginal, conditional, or population-level data) affect the identifiability of total and direct causal effects in the algebraic framework?

Key findings

• The paper establishes that causal models can be redefined using a finite set of unfolding events and a partial order, without requiring a graphical structure.
• Causal interventions are defined as projections of the joint distribution under the uncontrolled system, preserving the analogy to do-calculus in CBNs.
• The total causal effect of banning movie viewing on fighting behavior is identifiable only when combining data from experiments measuring testosterone levels and viewing behavior, with identification requiring division by the probability of exposure.
• The direct effect of low testosterone on fighting, represented as $ \widehat{\widehat{p}}(1,1,1,1) = \pi_2(1)\pi_4(1|1,1) $, is identifiable from a combination of experiments measuring marginal probabilities and conditional distributions.
• The framework allows for the systematic analysis of identifiability and feasibility using algebraic geometry, with the potential to exploit symmetries not visible in graphical representations.
• While the movie example uses simple algebraic operations, the method generalizes to complex models where computational algebra systems are required, though current software limitations persist due to inequality constraints and high-dimensional probability spaces.