Emerging Directions in Machine Learning

"Correlation does not imply causation" is perhaps the most frequently repeated mantra in statistics and data science. Yet for all its repetition, machine learning—the field that has revolutionized AI over the past decade—is fundamentally built on learning correlations. Neural networks, gradient boosting, random forests, and virtually every other mainstream ML algorithm are designed to discover and exploit patterns in data, not to understand why those patterns exist.

This distinction matters profoundly. A model that learns ice cream sales and drowning deaths are correlated might predict drownings from ice cream sales (summer causes both). But this model is useless—even dangerous—for decision-making: banning ice cream won't prevent drownings. Understanding the underlying causal structure (summer → ice cream, summer → swimming → drowning) is essential for reasoning about interventions.

Causal Machine Learning represents a paradigm shift: extending ML from learning 'what is associated with what' to understanding 'what causes what' and 'what would happen if we acted differently.' This shift is essential for building AI systems that can plan, reason about counterfactuals, transfer knowledge across domains, and provide explanations that support decision-making.

Judea Pearl, one of the founding figures of causal inference, articulated the Ladder of Causation—a hierarchy of cognitive capabilities that distinguishes different levels of causal understanding. This framework provides essential context for understanding what causal ML aims to achieve and why it represents such a significant extension beyond standard ML.

Rung 1: Association (Seeing)

The first rung involves observational queries: "What is the probability of Y given that we observe X?" This is the domain of traditional statistics and machine learning.

• Example question: What is the probability a patient has diabetes given they have high blood sugar?
• Mathematical form: P(Y | X)
• What it captures: Statistical correlations and conditional probabilities in observed data

Standard machine learning lives on this rung. All of supervised learning—classification, regression, sequence modeling—fundamentally learns to predict one variable given observations of others. This is powerful but limited: observed correlations can arise from many different causal structures.

Rung 2: Intervention (Doing)

The second rung involves interventional queries: "What would happen if we actively set X to some value?" This differs fundamentally from observation—it asks about the effects of actions, not passive associations.

• Example question: What would happen to a patient's blood sugar if we give them insulin?
• Mathematical form: P(Y | do(X = x))
• What it captures: The causal effect of actively manipulating a variable

Note the crucial difference: observing that someone has low blood sugar tells us something about their overall health (maybe they're diabetic and overmedicated). But giving someone insulin and observing the result isolates the causal effect of insulin specifically.

The 'do' operator, introduced by Pearl, represents this distinction mathematically. P(Y | do(X)) is fundamentally different from P(Y | X), and confusing them leads to fallacious causal reasoning.

Rung 3: Counterfactuals (Imagining)

The third rung involves counterfactual queries: "What would have happened if things had been different?" These questions reason about alternative realities that didn't occur.

• Example question: Would this patient have recovered if we had given them treatment A instead of treatment B?
• Mathematical form: P(Y_x | X = x', Y = y) — the probability that Y would have been different under intervention x, given that we observed X = x' and Y = y
• What it captures: Reasoning about specific individuals in alternative scenarios

Counterfactual reasoning is ubiquitous in human cognition. Legal and moral judgments often hinge on counterfactuals: 'Would the death have occurred had the defendant acted differently?' Effective learning from experience requires counterfactual reasoning: 'What would have happened if I had studied more for that exam?'

Counterfactual queries are strictly more powerful than interventional queries, which are strictly more powerful than associational queries. A model that can answer counterfactual questions can answer interventional and associational questions, but not vice versa.

The Data Wall

A crucial insight from the ladder of causation is that you cannot ascend the ladder through data alone. No amount of observational data, collected through passive observation, can answer interventional or counterfactual questions without additional assumptions about the causal structure.

This is why randomized controlled trials are the 'gold standard' for causal inference—they physically implement the do operator through random assignment, bypassing confounding. But when experiments are impossible, expensive, or unethical, we need causal reasoning frameworks that combine data with explicit causal assumptions.

The mathematical foundation for modern causal reasoning is the Structural Causal Model (SCM), which provides a rigorous framework for representing and reasoning about causal relationships. Understanding SCMs is essential for any serious engagement with causal ML.

Definition of an SCM

A Structural Causal Model M consists of:

1. Endogenous variables (V): The variables within the system we're modeling, whose values are determined by other variables in the model.

2. Exogenous variables (U): Background or external variables that influence the system but whose values are determined outside the model. These represent unexplained variation.

3. Structural equations (F): A set of functions f_i that determine each endogenous variable V_i based on its direct causes (parents) and exogenous factors: V_i = f_i(Parents(V_i), U_i)

4. Distribution over U: A probability distribution P(U) over the exogenous variables.

The structural equations encode the causal mechanism by which each variable is generated from its causes. Critically, these are asymmetric—they specify how causes produce effects, not just correlations.

From SCMs to Graphs

Every SCM implies a directed acyclic graph (DAG) G, called the causal graph, where:

• Nodes represent endogenous variables
• Directed edges X → Y indicate that X is a direct cause of Y (appears in Y's structural equation)

The graph provides a visual representation of the causal structure and enables powerful graphical criteria for reasoning about causal effects, such as d-separation for determining conditional independence and the backdoor/frontdoor criteria for identifying causal effects from observational data.

The Do-Calculus

Pearl's do-calculus provides a complete set of inference rules for manipulating expressions involving the do operator. Given a causal graph, these rules determine when and how interventional quantities P(Y | do(X)) can be computed from observational data P(V).

The three rules of do-calculus are:

1. Insertion/deletion of observations: P(Y | do(X), Z, W) = P(Y | do(X), W) if Y ⊥⊥ Z | X, W in the graph with incoming edges to X removed.

2. Action/observation exchange: P(Y | do(X), do(Z), W) = P(Y | do(X), Z, W) under specific graphical conditions.

3. Insertion/deletion of actions: P(Y | do(X), do(Z), W) = P(Y | do(X), W) under specific graphical conditions.

A causal effect P(Y | do(X)) is said to be identifiable if it can be computed from observational data using do-calculus. Remarkably, do-calculus is complete: if an effect is identifiable, the do-calculus rules suffice to derive the formula; if it's not derivable, no method can identify it from observational data alone.

Practical Identification: Backdoor and Frontdoor Criteria

Two important special cases simplify identification in common scenarios:

Backdoor Criterion: A set Z satisfies the backdoor criterion relative to (X, Y) if:

1. Z blocks all backdoor paths from X to Y (paths that enter X through arrows pointing into X)
2. Z contains no descendants of X

If Z satisfies the backdoor criterion: P(Y | do(X)) = Σ_z P(Y | X, Z) P(Z)

This is the adjustment formula—we can estimate causal effects by adjusting for confounders.

Frontdoor Criterion: When direct observation of confounders is impossible, the frontdoor criterion provides an alternative identification strategy through intermediate variables that mediate the effect of X on Y.

An alternative framework for causal inference, developed primarily by Donald Rubin and coworkers, centers on potential outcomes. This framework is particularly prominent in statistics, epidemiology, and economics. Understanding both SCMs and potential outcomes is essential, as each offers complementary insights.

The Potential Outcomes Framework

For each unit i (e.g., a patient) and each possible treatment value x, we define a potential outcome Y_i(x)—the outcome that would be observed if unit i received treatment x.

For a binary treatment (treated/untreated):

• Y_i(1): The outcome if unit i receives treatment
• Y_i(0): The outcome if unit i does not receive treatment

The Individual Treatment Effect (ITE) for unit i is: ITE_i = Y_i(1) - Y_i(0)

This is the causal effect of treatment for individual i. However, we face the fundamental problem of causal inference: for any individual, we observe only one potential outcome—the one corresponding to the treatment they actually received. The counterfactual outcome is never observed.

Estimating Average Effects

Since individual treatment effects are unobservable, we typically estimate average effects:

Average Treatment Effect (ATE): ATE = E[Y(1) - Y(0)] = E[Y(1)] - E[Y(0)]

Average Treatment Effect on the Treated (ATT): ATT = E[Y(1) - Y(0) | T = 1]

Conditional Average Treatment Effect (CATE): CATE(x) = E[Y(1) - Y(0) | X = x]

Estimating CATE—how treatment effects vary with individual characteristics—is a major focus of causal ML.

The Ignorability Assumption

For valid causal inference from observational data, we typically require conditional ignorability (also called unconfoundedness or selection on observables):

Y(0), Y(1) ⊥⊥ T | X

This states that, conditional on observed covariates X, treatment assignment T is independent of potential outcomes. In other words, there are no unmeasured confounders—all variables that influence both treatment selection and outcomes are observed.

Ignorability is a strong assumption and is generally untestable from data alone. Whether it's plausible depends on domain knowledge about the data-generating process.

Connection to SCMs

The potential outcomes framework and SCMs are formally connected:

• Each potential outcome Y_i(x) corresponds to the counterfactual Y_x in the SCM framework
• Ignorability corresponds to the backdoor criterion being satisfied by observed covariates
• Both frameworks agree on identifiability conditions and effect estimation formulas

Pearl's SCM framework is more general (it represents the complete causal model), while potential outcomes focus specifically on treatment effect estimation. In practice, researchers often use potential outcomes notation for effect estimation while implicitly assuming an underlying SCM structure.

Propensity Scores

A key tool in potential outcomes analysis is the propensity score e(x) = P(T = 1 | X = x)—the probability of receiving treatment given observed covariates.

Rosenbaum and Rubin showed that if ignorability holds given X, it also holds given e(X) alone. This enables matching or weighting on a single scalar rather than high-dimensional covariates:

Inverse Propensity Weighting (IPW): ATE = E[Y · T / e(X)] - E[Y · (1-T) / (1-e(X))]

The intersection of causal inference and machine learning has produced a rich set of methods that leverage ML's powerful function approximation capabilities for causal effect estimation. These methods are particularly valuable for estimating heterogeneous treatment effects—how causal effects vary across individuals.

Meta-Learners

Meta-learners are flexible frameworks that use arbitrary ML models as base learners for CATE estimation:

T-Learner (Two-Model):

1. Train model μ₀(x) on control group: E[Y | X, T=0]
2. Train model μ₁(x) on treated group: E[Y | X, T=1]
3. Estimate: CATE(x) = μ₁(x) - μ₀(x)

Simple but can be biased when treatment assignment is imbalanced or when there's limited overlap.

S-Learner (Single-Model):

1. Train single model μ(x, t) on all data: E[Y | X, T]
2. Estimate: CATE(x) = μ(x, 1) - μ(x, 0)

Can struggle to capture treatment effect when effect is small relative to baseline variation.

X-Learner:

1. Train T-learner models μ₀, μ₁
2. Impute treatment effects in each group using cross-predictions
3. Weight estimates based on propensity scores

More robust than T-learner, especially with treatment imbalance.

DR-Learner (Doubly Robust): Combines outcome modeling with propensity weighting to achieve robustness—estimates are consistent if either the outcome model or the propensity model is correctly specified.

Causal Forests

Causal forests (Wager & Athey, 2018) adapt random forests specifically for causal effect estimation. Key innovations include:

1. Honesty: Trees are built using one subsample and predictions are made using another, avoiding overfitting to the outcome.

2. Orthogonalization: Residualized treatment and outcomes are used to focus the forest on the treatment effect rather than the baseline outcome.

3. Variance Estimation: The method provides valid confidence intervals for CATE estimates.

Causal forests have become a standard tool for discovering heterogeneous treatment effects and are implemented in the popular grf (generalized random forests) package.

CATE Neural Networks

Deep learning approaches to CATE estimation include:

• DragonNet: Uses a three-headed neural network architecture that jointly learns representations, propensity scores, and conditional outcomes, with a regularization term that encourages representations to be prognostic of treatment.

• CEVAE: Causal Effect Variational Autoencoder uses variational inference to model latent confounders when observational data is subject to hidden confounding.

• TARNet: Treatment-Agnostic Representation Network learns a shared representation layer with separate treatment-specific heads.

These methods leverage neural networks' ability to learn complex functions while incorporating causal inference principles into the architecture and training objectives.

The methods discussed so far assume the causal graph is known. But where does this knowledge come from? Causal discovery addresses the problem of learning causal structure from data—inferring the causal graph rather than assuming it.

The Identifiability Challenge

Causal discovery is fundamentally limited by what can be learned from observational data. Multiple causal graphs can be consistent with the same observational distribution. Specifically, two graphs that share the same Markov equivalence class—encoding the same conditional independencies—cannot be distinguished from observational data alone.

For example, A → B and A ← B both imply P(A,B) = P(A)P(B|A) = P(B)P(A|B). Without interventional data or additional assumptions, we cannot determine the direction of causation.

Constraint-Based Methods

Constraint-based algorithms discover causal structure by testing conditional independencies in data:

PC Algorithm (Peter-Clark):

1. Start with a complete undirected graph
2. Remove edges between conditionally independent variables
3. Orient edges based on v-structures (colliders) and acyclicity constraints

FCI Algorithm (Fast Causal Inference): Extends PC to handle hidden confounders, producing a Partial Ancestral Graph (PAG) that represents the equivalence class of possible causal structures.

Constraint-based methods are nonparametric—they don't assume specific functional forms. However, they're sensitive to errors in conditional independence testing, especially in high-dimensional settings.

Score-Based Methods

Score-based methods search over possible graphs to optimize a score function:

BIC/BDe Scoring: Penalized likelihood scores that balance fit to data against model complexity.

GES (Greedy Equivalence Search): Searches the space of Markov equivalence classes using edge additions and deletions, optimizing a decomposable score.

Score-based methods can be more robust to individual independence test errors but face the challenge of searching over a super-exponentially large space of possible graphs.

Functional Causal Models and Identifiability

Remarkably, under certain functional assumptions, the causal direction becomes identifiable:

Linear Non-Gaussian Acyclic Models (LiNGAM): If relationships are linear and noise is non-Gaussian, the causal direction is identifiable. The key insight is that for X → Y: Y = αX + ε with non-Gaussian X and ε, the residual in the wrong direction (regressing X on Y) is not independent of its predictor, while the residual in the correct direction is.

Additive Noise Models: More generally, if Y = f(X) + ε with independent noise, identifiability holds under various conditions on f and the noise distribution (e.g., non-Gaussian noise, or non-linear f).

Post-Nonlinear Models: Extend additive noise models to the form Y = g(f(X) + ε), allowing an additional output transformation.

Causal Discovery with Neural Networks

Recent approaches use neural networks for causal discovery:

• DAG-GNN, RL-BIC: Frame structure learning as a continuous optimization problem using acyclicity constraints and differentiable structure learning.

• NOTEARS: Reformulates the discrete acyclicity constraint as a differentiable equality constraint, enabling gradient-based DAG learning.

• DiffAN: Combines neural additive models with asymmetry-based causal direction testing.

These methods can scale to larger variable sets than traditional algorithms, though they face challenges with consistency guarantees and handling of latent confounders.

Interventional Data for Discovery

When interventional data is available, causal discovery becomes more powerful. Intervening on a variable breaks incoming edges in the graph, providing information about causal direction that observational data cannot.

Active learning for causal discovery selects which interventions to perform to maximally disambiguate the underlying causal structure. This is especially relevant in biology and other experimental sciences where interventions are possible but costly.

Standard causal inference assumes we observe the relevant variables. But in many modern ML applications, we work with high-dimensional raw data (images, text, audio) where the underlying causal variables are latent. Causal representation learning addresses how to discover and learn representations of these latent causal variables from raw observations.

The Problem

Consider a robot learning from images of a scene. The image pixels are the observations, but the underlying causal variables are things like:

• Object positions (ball is on table)
• Object properties (ball is red, small)
• Agent actions (robot arm moved)
• Physical laws (ball falls when unsupported)

The causal relationships exist at the level of these latent variables, not at the pixel level. A representation that explicitly captures these variables would enable causal reasoning, transfer learning, and compositional generalization.

Identifiability of Latent Causal Variables

A fundamental question is whether latent causal variables can be uniquely recovered (up to appropriate equivalence) from observations. Recent theoretical work has made significant progress:

Independent Component Analysis (ICA) Perspective: Classical ICA shows that independent non-Gaussian sources can be recovered from linear mixtures. Nonlinear ICA is generally unidentifiable, but becomes identifiable with additional information:

• Temporal structure (sources have temporal dependencies)
• Auxiliary information (class labels, multi-view data)
• Known intervention structure

Causal Extensions: Schölkopf and collaborators have developed identifiability theory for nonlinear causal representations under assumptions like:

• Known causal graph over latent variables
• Access to interventional data
• Multi-environment data with known environment-variable relationships

Practical Approaches

Variational Autoencoders for Causal Latents: Extensions of VAEs that encourage disentangled or causally-structured latent spaces:

• β-VAE: Higher regularization pressure on the latent space encourages more factorized representations
• CausalVAE: Incorporates a causal graph prior over latent variables
• DEAR: Disentangled generative causal representation learning with explicit intervention mechanisms

Contrastive Learning for Causality: Contrastive methods that compare views under causal equivalence:

• Invariant Risk Minimization (IRM): Learn representations that support classifiers optimal across different environments
• Causal contrastive learning: Define positive pairs based on causal rather than perceptual similarity

Object-Centric Learning: Methods that decompose scenes into discrete object representations, enabling compositional understanding:

• MONet, IODINE, Slot Attention: Learn to segment scenes into object slots
• C-SWM, CODA: Learn object representations with dynamics for causal world models

The Big Picture: Toward Causal Foundation Models

A grand vision for causal representation learning is developing foundation models that:

1. Represent the world in terms of causally relevant variables
2. Learn generalizable causal relationships among these variables
3. Support counterfactual reasoning: 'What would this scene look like if the ball were removed?'
4. Enable robust transfer: learned relationships apply in new domains with the same underlying structure

This remains largely aspirational, but progress in identifiability theory, multi-environment learning, and compositional representation learning is moving toward this vision.

Causal ML methods are increasingly deployed in high-stakes domains where understanding 'what would happen if we acted differently' is essential. These applications demonstrate the practical value of moving beyond correlation to causation.

Personalized Medicine and Treatment Effect Estimation

Perhaps the most developed application domain, where causal ML directly improves patient outcomes:

• Treatment Heterogeneity: Identifying which patients benefit most from specific treatments. A drug might have no average effect but substantial benefit for a subpopulation—causal ML can identify this subpopulation.

• Optimal Treatment Rules: Learning policies that recommend treatments based on individual characteristics, maximizing expected outcomes.

• Observational Studies: Estimating causal effects from electronic health records when randomized trials are impractical, ethical concerns apply, or real-world generalizability is needed.

Tech Industry: A/B Testing and Beyond

Tech companies apply causal inference at massive scale:

• Heterogeneous Treatment Effects: Moving beyond average effects to understand how product changes affect different user segments.

• Long-term Effects: Estimating long-term causal impacts when only short-term outcomes are observed.

• Network Effects: Accounting for interference when users influence each other, violating the standard assumption of independent observations.

• Continuous Experimentation: Designing efficient experimental systems that learn optimal policies while minimizing regret.

We've explored the rich landscape of causal machine learning, from foundational concepts to cutting-edge research. Let's consolidate the key insights:

What's Next:

Having explored causal ML's approach to understanding 'why,' we'll next examine World Models—systems that learn internal simulators of their environment to enable planning, imagination, and transfer. World models represent another crucial step toward AI systems that truly understand the world rather than merely pattern-matching on surface features.