PaperlayerView exact revision on arXiv

Source-linked AI summary

Causal Machine Learning: A Survey and Open Problems

Jean Kaddour, Aengus Lynch, Qi Liu, Matt J. Kusner, Ricardo Silva

arXiv:2206.15475v3cs.LGstat.ME

TL;DR

Machine-learning models can struggle when deployment data differ from training data, and hidden confounding can make spurious associations unreliable. This survey compares causal machine-learning methods across problem areas and concludes that causal supervised learning improves predictive generalization through invariant features or mechanisms.

  • Problem

    Machine-learning models can lose generalization under distribution shift, while hidden confounding can make spurious training associations unreliable in deployment.

    Source: Introduction · p. 6, §2.5 · p. 26

  • Method

    The survey compares causal supervised learning and causal explanations, and reviews modality-specific applications and principles.

    Source: Conclusion · pp. 194–195

  • Results

    Causal supervised learning improves predictive generalization by learning invariant features or mechanisms.

    Source: Conclusion · p. 194

  • Takeaways & Limitations

    Modality-specific principles offer opportunities to improve computer vision, natural language processing, and graph representation learning.

    Source: Conclusion · p. 195

    Causal methods may rely on assumptions that fail in practice or cannot be tested with finite data, especially when perturbations are scarce or confounders unobserved.

    Source: §3.3.4 · p. 54

Abstract

from arXiv · show

Causal Machine Learning (CausalML) is an umbrella term for machine learning methods that formalize the data-generation process as a structural causal model (SCM). This perspective enables us to reason about the effects of changes to this process (interventions) and what would have happened in hindsight (counterfactuals). We categorize work in CausalML into five groups according to the problems they address: (1) causal supervised learning, (2) causal generative modeling, (3) causal explanations, (4) causal fairness, and (5) causal reinforcement learning. We systematically compare the methods in each category and point out open problems. Further, we review data-modality-specific applications in computer vision, natural language processing, and graph representation learning. Finally, we provide an overview of causal benchmarks and a critical discussion of the state of this nascent field, including recommendations for future work.

Introduction

The survey uses structural causal models (SCMs) to frame how causal knowledge can address open machine-learning problems, and organizes CausalML methods while reviewing applications and benchmarks.

Introduction

  • Real-world ML faces distribution-shift failures [1], weak generative control, unfair bias,, abstract interpretability, and unstable RL transfer [6], motivating causal formalisms, [8],, [10], [11].The survey focuses on SCMs, which model interventions on the data-generating process and counterfactual consequences in hindsight.Source: Introduction · pp. 6–7

Contributions

  • The survey taxonomizes CausalML into causal supervised learning, causal generative modeling, causal explanations, causal fairness, and causal reinforcement learning.It also provides a self-contained introduction to causality without assuming prior knowledge.Source: Contributions · p. 7, causal explanations · p. 7

causal explanations

  • For each problem class, the survey compares existing methods and identifies avenues for future work.The survey addresses a lack of clear categorization, in part because CausalML assumptions can be unfamiliar, difficult to relate across setups, and untestable.Source: causal explanations · p. 7, Introduction · p. 7

causal reinforcement learning

  • The survey reviews causal applications in computer vision, natural language processing, and graph representation learning, alongside causal benchmarks.Source: causal reinforcement learning · p. 7

4 Introduction

The survey discusses CausalML’s benefits compared with non-causal ML, the challenges facing the field, and the costs of using causal techniques.

4 Introduction

  • Chapter 10 covers CausalML’s benefits over non-causal ML, the field’s challenges, and the costs of using causal techniques.The survey frames these as the good, the bad, and the ugly, respectively.Source: §4 · p. 8

Causality: A Minimal Introduction

This section introduces causal models as a framework for representing causal mechanisms and reasoning about observations, interventions, and counterfactuals. It covers causal graphs and conditional factorization, confounding, causal representation learning, estimand identification, and causal influence, including its definition using KL divergence and an interventional distribution after removing an edge.

2.1. Bayesian Networks

  • Graphical factorization reduces modeling complexity: the farmer example replaces a full joint factorization with p(t)p(d | t)p(r | t)p(s | d, r).The naive factorization can require exponentially many parameters as the number of variables grows.Source: §2.1 · pp. 11–12

2.1 Bayesian Networks

  • Causal DAGs encode assumptions about relationships among variables and factor joint distributions into smaller conditionals based on graph parents.Under the Markov condition, each node is independent of its non-descendants given its parents, enabling the joint distribution to factorize over parent-conditioned variables.Source: §2.1 · pp. 12–13

2.2.1 Interventions

  • An observational distribution p(s | d) conditions on cases with D = d, whereas p(s | do(d)) describes the population after D is actively set to d.The intervention removes D’s incoming edge from T, so the season no longer determines the sprinkler setting; valid causal BNs therefore require assumptions about how interventions change mechanisms.Source: §2.2.1 · pp. 14–16
  • Equivalent conditional-independence constraints can arise from DAGs with reversed edges, so a valid probabilistic factorization alone does not establish causal direction.For example, the section notes that a graph with slipperiness causing rain can still factorize the joint distribution for statistical inference, despite conflicting with common-sense causal assumptions.Source: §2.2.1 · pp. 16–17

Examples of Interventions

  • Examples distinguish observation from intervention: breaking a barometer changes its reading without changing boiler pressure, while treatment interventions compare outcomes across all patients rather than only treated patients.Thus p(y | do(x)) differs from p(y | x), and p(y | do(t)) represents setting treatment t across patients with medical-history features x.Source: Examples of Interventions · p. 17

2.2.2 Counterfactuals

  • A counterfactual combines observed facts with an unrealized intervention while holding other environmental conditions fixed, enabling comparison of outcomes that differ in one aspect.The section distinguishes the actual outcome s0 from the counterfactual outcome under D = d′, since conditioning on both actual and hypothetical slipperiness creates a clash.Source: §2.2.2 · p. 18

Examples of Counterfactuals

  • Counterfactuals ask what would have happened retrospectively for an observed case; unlike interventions, they preserve the case’s features and outcome, and sit above interventions and observations in the causal hierarchy.For image editing, counterfactual sampling aims to preserve the object while changing its background, unlike interventional sampling, which may produce diverse images.Source: Examples of Counterfactuals · p. 19, §2.2.3 · p. 19, Table 2.1 · p. 20

2.3 Structural Causal Models

  • Structural Causal Models (SCMs) use deterministic structural equations with unobserved noise to represent causal relations, induce a DAG, and entail conditional independence of each variable from its non-descendants given its parents.SCMs enable counterfactual analysis that causal Bayesian networks cannot construct.Source: §2.3 · pp. 20–21

2.3.1 Interventions

  • In an SCM, interventions replace the targeted structural assignments while retaining the prior noise distribution; counterfactuals instead use an individual’s observed data to infer exogenous noise.Counterfactual answers are individual-level, whereas interventional distributions represent population-level effects.Source: §2.3.1 · pp. 21–22

2.3.2 Counterfactual Inference

  • Counterfactual inference proceeds by abduction of noise compatible with observations, action through replacing the intervened equations, and prediction using the modified SCM.The counterfactual SCM replaces the prior noise distribution p(ϵ) with the posterior p(ϵ | x).Source: §2.3.2 · p. 22, §2.4 · p. 23

2.3.3 Independent Mechanisms

  • The Principle of Independent Mechanisms makes interventions local: targeted variables’ mechanisms are replaced, while mechanisms for untargeted variables remain invariant, allowing one graph to encode multiple interventional distributions.For a targeted variable, the intervention sets its conditional mechanism to the imposed value.Source: §2.3.3 · p. 23

2.4 Causal Representation Learning

  • Causal representation learning treats representations as instances of latent causal variables, but learning their full SCM is difficult without extensive supervision or domain knowledge.Its components are causal feature learning, causal graph discovery, and causal mechanism learning.Source: §2.4 · p. 24

2.5. Spurious Relationships due to Confounding

  • Hidden confounding can make models rely on spurious training associations that may fail in production; ImageNet classifiers are more likely to misclassify objects shown in different backgrounds.The example describes bird images whose backgrounds reflect an unobserved environment variable.Source: §2.5 · p. 26

2.6 Causal Estimand Identification

  • A causal estimand is identifiable when it can be computed from a statistical quantity; with no hidden confounders and a known graph, backdoor adjustment can identify an average treatment effect [13].Causal influence is another quantity of interest; Janzing et al. conclude that KL-divergence satisfies natural desiderata for measuring it.Source: §2.6 · p. 27, §2.7 · p. 28

2.7 Causal Influence

  • Causal influence from node k to node l in an SCM is defined using KL divergence and the interventional distribution after removing the edge from k to l.The supplied passage does not include the full formula.Source: §2.7 · p. 28

Causal Supervised Learning

This section surveys causal supervised learning methods for domain-robust prediction, including approaches that address confounding, exploit causal invariances, and learn representations for shifted or novel domains. It also discusses limitations in identifying and testing causal assumptions, benchmark gaps, and directions involving controlled invariance benchmarks, adversarial robustness, meta-learning, and richer supervision signals.

Causal Supervised Learning

  • When i.i.d. assumptions fail, spurious feature-label associations can undermine prediction; causal methods instead target robust features or mechanisms across interventional environments.Interventions and counterfactual claims still depend on assumptions about the data-generating process, such as no hidden confounding.Source: Causal Supervised Learning · pp. 30–31

3.1 Invariant Feature Learning

  • Invariant feature learning (IFL) seeks content features comprising the causal parents of Y, whose relationship to Y remains predictive across interventions unless Y itself is intervened upon.The Style and Content Decomposition separates latent content from style, but assumes no hidden confounders between those groups.Source: §3.1 · pp. 33–34

Deconfounding Data

  • Data augmentation can intervene on style to break spurious associations with labels; using observational data, it can also simulate new environments and enable automatic selection of transformations for improved domain generalization.Source: Deconfounding Data · pp. 35–36

Deconfounding feature maps

  • Causal transportability [52] splits pre-trained embeddings into style and content variables, then marginalizes style contributions to learn predictions using causal parents of Y.The method assumes disjoint style and content embeddings and a confounded path from style through content to Y; training uses corrupted inputs, while evaluation performs marginalization.Source: Deconfounding feature maps · pp. 37–38

Deconfounding models during training

  • Counterfactual invariance [53] requires a predictor to give the same output for a sample and its counterfactual under changes to a specified attribute A.Practitioners must identify A and the causal direction between X and Y to choose the corresponding regularizer.Source: Deconfounding models during training · pp. 38–39
  • Asymmetry learning encodes specified transformation invariances in the training objective; composing such transformations models distribution shifts, but generalization requires specifying which remain invariant.Equivalence relations define transformations that preserve class membership.Source: Deconfounding models during training · p. 40
  • Contrastive self-supervised augmentations simulating style interventions can isolate invariant content features; ReLIC uses proxy-task prediction and an augmentation-invariance KL term.Self-supervised approaches may be reconstructive or discriminative and can use unlabeled data for downstream tasks.Source: Deconfounding models during training · pp. 40–41

Deconfounding Predictions Post-Hoc

Feature Maps with Invariant Classifiers

  • Invariant Risk Minimization learns a feature map and linear predictor that minimize worst-case environment-specific risk, but can underperform ERM when causal relations are nonlinear or environments lack sufficient heterogeneity [37].IRM-like methods have OOD guarantees for linear regression but not necessarily linear classification; information-bottleneck constraints resolve some issues under distribution shifts.Source: §3.1.1 · p. 43, Feature Maps with Invariant Classifiers · pp. 43–44
  • Causal-Rep uses probability of necessity and sufficiency, ISR-Mean generalizes with ds + 1 environments, and ISR-Cov reduces the requirement to O(1) [30, 74].REx seeks robustness across affine combinations of training distributions; Jiang et al. [73] address anti-causal shifts, where covariates are caused by the outcome and naive invariant prediction is vacuous.Source: Feature Maps with Invariant Classifiers · pp. 45–46

Object Matching

SCD with priors

  • LaCIM uses environment-specific encoders and priors with shared decoders, CSG trades those priors for a single training environment while assuming independent style and content, and iCaRL learns causal parents in a three-stage procedure.iCaRL first learns disentangled latents, then uses the PC algorithm to identify causal parents of Y, and finally trains a classifier with ERM.Source: SCD with priors · p. 47

3.2.1 Independent Invariant Mechanisms as Separate Networks

  • Invariant mechanism learning identifies mappings for interventional distributions rather than confounder feature maps; RIMs use competing mechanisms for sequential data [83, 84].RIMs use attention to select and connect mechanisms, improving long-dormant information retention and multi-object trajectory prediction over LSTMs and transformers; a fast/slow meta-learning variant improves over LSTM and RIMs on meta-learning benchmarks.Source: §3.2 · p. 49, §3.2.1 · p. 50

3.2.2 Invariant Mechanisms as Mappings Between Domains

  • Domain-mapping methods infer disentangled mechanisms using proxies for unobserved confounders to map test inputs counterfactually to source inputs, or use nonlinear ICA to generate target-domain samples from invariant mappings.The generated pseudo-samples can be used to train a standard supervised learning algorithm.Source: §3.2.2 · p. 50, §3.2 · p. 51

3.3.1 Lack of Targeted Benchmarks for Invariance Learning

  • Invariance methods often rely on separate toy experiments rather than standardized test beds; ColoredMNIST evaluates models under changed color-digit associations between training and test data [37].This benchmarking gap makes it difficult for practitioners to determine which approach best fits their problems.Source: §3.3.1 · p. 52
  • Synthetic invariance benchmarks can control spurious correlations while preserving photo-realism, with difficulty levels and train/test shifts; text-to-image models enable large-scale datasets [36].Earlier proposals use CelebA’s black-hair/necklace association and compare generalization under changes to a specified representation dimension.Source: §3.3.1 · p. 52

3.3.2 Connections between Invariance Learning, Adversarial Robustness and Meta-learning

  • Adversarial robustness can be viewed as invariance to perturbations, while meta-learning’s shared task structure is task-invariant; causal approaches to both remain underexplored.The authors suggest exploring interventions on task variables to learn better task-invariant representations.Source: §3.3.2 · p. 53

3.3.3 Exploiting Additional Supervision Signals

  • Future work can combine augmentation-generated environments with multi-environment algorithms and exploit high-dimensional environment descriptors rather than environment indices.Examples include using an aerial landscape that causes both data and label, or robotics task descriptors such as one image per environment.Source: §3.3.3 · pp. 53–54

3.3.4 Limitations in Causal Identifiability and Unverifiable Assumptions

  • Causal invariance methods may be unlearnable or untestable when decomposition or independence assumptions fail, data variation is insufficient, or confounders obscure causal paths.The authors call for clarifying which assumptions are essential and designing benchmarks or protocols to test them.Source: §3.3.4 · p. 54

Causal Generative Modeling

This section surveys causal generative modeling methods that use interventional and counterfactual distributions for controllable generation and editing, including intervention-based learning of latent causal models. It also identifies open challenges in choosing causal abstractions, scaling causal graphs, and evaluating counterfactual data augmentation.

4.1 Structural Assignment Learning

  • Learning causal mechanisms depends on assumptions that can be difficult to verify; hidden confounders and large graphs can undermine structural assignments.Structural assignment learning typically assumes a known or approximated DAG, and unobserved confounders become hidden causes when omitted from the labeled variables.Source: Causal Generative Modeling · p. 55, §4.1 · pp. 57–58

Causal Generative Modeling

  • Interventional generation can produce attribute combinations absent from training data, while counterfactual editing changes specified attributes and preserves unrelated ones.These approaches estimate p(x | do(a)) for generation and p(x_A→a | x) for editing.Source: Causal Generative Modeling · pp. 55–56

4.1.1 Independently Learned Structural Assignments

  • DeepSCM learns assignments under a supplied DAG and abducts sample noise for counterfactuals, while VACA [100] jointly learns nonlinear assignments but struggles as graph depth grows.DeepSCM has been applied to graphs of order five; VACA needs decoder depth at least the longest causal path length minus one, while deep GNNs risk over-smoothing.Source: §4.1.1 · pp. 58–59, §4.1.2 · pp. 59–60

4.1.3 Group Structural Assignments

  • DCEVAE [101] segments a causal graph around intervention targets to support higher-order counterfactual sampling, improving on causal-agnostic models but producing often-blurry images.The model uses observed concept labels and the intervention targets’ causal descendants to generate counterfactual samples.Source: §4.1.3 · p. 60

4.1.4 Structural Assignments with a GAN

  • DEAR [102] uses a GAN to learn structural assignments among latent variables, but requires causal ordering and sample-level labels for the variables of interest.It initializes a supergraph in the latent prior and identifies the adjacency structure during training.Source: §4.1.4 · pp. 61–62

4.2 Diffusion-Based Counterfactual Estimation

  • Diff-SCM [108] uses anti-causal classifier guidance to simulate label interventions, generating counterfactual samples with changed classifications but some high-level similarities.Forward diffusion represents the exogenous factors, and a parameter balances classifier guidance against the score-matching objective.Source: §4.2 · p. 62

4.3 Disentangled Mechanisms via Domain Knowledge and 4.4 Disentangled Mechanisms via Intervention Data

  • CGN composes shape, texture, and background mechanisms assuming independent factors; Ada-GVAE instead uses paired pre/post-intervention data and the number of changed latent variables.CGN shares noise and attributes across mechanisms for counterfactuals, whereas Ada-GVAE identifies intervened factors by comparing latent posteriors across sample pairs.Source: §4.3 · p. 63, §4.4 · pp. 63–64

4.5 Causal Disentanglement

  • Causal disentanglement avoids requiring a complete graph but needs knowledge of target variables; CausalVAE uses labels, while ELCM [112] discovers graphs without requiring intervention counts.CausalVAE learns linear structural assignments, whereas VACA allows nonlinear ones.Source: §4.5 · p. 65, §4.5.1 · pp. 65–66, §4.5.2 · p. 66

4.5. Causal Disentanglement

  • ILCM maps pre- and post-intervention data to exogenous noise without requiring a specified graph or mechanisms, then recovers explicit generative variables and a graph through solution functions and intervention-based discovery.It requires the latent dimension K and assumes each intervention changes only one noise value.Source: §4.5.2 · p. 66, §4.5 · p. 67

Exploit sequential data with intervention labels

  • CITRIS assumes no instantaneous effects and learns latent variables and a transition prior from intervention-labeled sequences, grouping dimensions of the same causal factor when needed.Its prior is trained with an ELBO objective and can represent a multidimensional position variable rather than separate axis variables.Source: Exploit sequential data with… · p. 67
  • CITRIS factorizes latents into causal variables and exogenous noise, then uses transition-prior dependencies to identify parents and the causal graph.It learns an assignment function mapping latent dimensions to causal variables.Source: Exploit sequential data with… · pp. 67–68
  • Because instantaneous effects can defeat CITRIS, Lippe et al. [114] propose iCITRIS, which assumes perfect interventions and learns graph and generative-model parameters by bi-level optimization.Unlike CITRIS, iCITRIS augments the transition prior with instantaneous graph structure.Source: Exploit sequential data with… · p. 68

4.6.1 Different Levels of Abstraction

  • Causal representation learning is ill-posed without choosing an abstraction level; pixel- and object-level models may perform equally, but the latter is more efficient and interpretable, while choosing levels and supervisory signals remains open.The approaches reviewed require specifications such as variable labels or latent dimension and paired pre/post-intervention samples.Source: §4.6.1 · p. 69
  • Theoretical work shows that identifying causal representations requires inductive biases or supervision, while multiple environments and constrained interventions offer routes to identifiability,,, [120].Disentangled representations can permit impossible combinations of causal variables; intervention frameworks can incorporate relations that constrain what is physically possible.Source: §4.6.1 · pp. 69–70

4.6.2 Scaling Structural Assignment Learning

  • Reviewed structural-assignment methods generally do not demonstrate graphs larger than order 5, apart from amortized DCEVAE, and VACA's GNNs can struggle to scale, leaving graph scalability open.These methods either use relatively low-order graphs or amortize assignments around plausible interventions.Source: §4.6.2 · p. 70

4.6.3 Understanding Counterfactual Data Augmentation

  • Sauer et al. found counterfactual GAN augmentations improved classifier performance more than non-counterfactual samples; when and how much they improve generalization and their use in benchmarking remain open.One proposed benchmark compares a fixed predictor's test loss after training with counterfactual samples generated by different methods.Source: §4.6.3 · p. 71

Causal Explanations

This section surveys causal explanations, covering local, single-input, model-agnostic feature-attribution and contrastive methods rather than global or model-specific explanations. It also discusses recourse-aware explanations, their integration with feature attribution, and challenges involving scalability, changing distributions, security and privacy, and competing explanation properties.

Causal Explanations

  • The surveyed methods are local explanations of individual predictions from a single input and are model-agnostic; global and model-specific methods are outside scope.The section focuses on feature attribution and contrastive explanations.Source: Causal Explanations · p. 72

5.1 Feature Attribution Explanations

  • Causal feature attributions rank features by their marginal contribution; unlike associational scores that may credit correlated features, they can direct attention to causes such as income.In the loan example, the applicant may focus on increasing income rather than changing careers because of features spuriously correlated with income.Source: §5.1 · pp. 73–74

5.1.1 CXPlain

  • CXPlain trains separate supervised explanation models using causal influence measured as the decrease in predictive error, requiring N(D + 1) target-model evaluations for N samples and D features.Bootstrap ensembles of M explanation models provide confidence intervals for feature-importance estimates.Source: §5.1.1 · pp. 74–75

5.1.2 Generative Causal Explanations

  • Generative causal explanations learn disentangled latent factors and quantify their causal influence on classifier outputs, encouraging a small subset of factors to have large impact while representing the data distribution.In the color-classifier example, changing the color factor affects the classifier output, while changing the shape factor does not.Source: §5.1.2 · pp. 75–76

5.1.3 Asymmetric Shapley Values

  • Asymmetric Shapley values [140] relax symmetry so causal ordering can favor ancestors over descendants, yielding distal-cause explanations while incorporating partial causal knowledge without requiring full causal inference.Known ancestors’ attributions reflect their effects while descendants remain unspecified; descendant attributions then represent incremental effects.Source: §5.1.3 · pp. 76–77

5.1.4 Causal Shapley Values

  • Causal Shapley values [142] replace observational conditioning with intervention to explain total effects and decompose them into direct and indirect effects; they can also incorporate a partial causal order.Janzing et al. argue that observational conditional distributions in prior Shapley methods overlook causal structure.Source: §5.1.4 · p. 78

5.2 Contrastive Explanations

  • Contrastive explanations emphasize why an event occurred rather than another, often using counterfactual changes to an individual’s features that would produce a desired model prediction.They prune the space of possible causal factors, reducing communication burden and cognitive load.Source: §5.2 · p. 79, §5.2.1 · p. 79

5.2.2 Causal Algorithmic Recourse via Minimal Interventions

  • Causal algorithmic recourse seeks cost-aware future actions that respect feature dependencies; ARMI models these as minimal-cost interventions yielding a favorable structural counterfactual, which need not be the nearest counterfactual explanation.The intervention’s effect on the individual’s situation is used to assess whether the desired outcome is achieved.Source: §5.2.2 · p. 81, (ARMI) [149] · p. 82

(ARMI) [149]

  • For ARMI [149], decision-boundary optimization may miss action effects; Kügelgen et al. bound recourse counterfactual effects with unobserved confounding and arbitrary structural equations.Their approach is compatible with unobserved confounding and bounds the expected counterfactual effect of recourse actions.Source: (ARMI) [149] · p. 82

5.3.1 Unifying Feature Attribution and Explanations

  • Feature attribution and explanations complement one another by showing changes and invariants [133]; early integrations,, [160],, and counterfactual-difference attribution support this synthesis.Causal attribution scores can signal that changing a feature should not interfere with the desired outcome.Source: §5.3.1 · p. 83

5.3.2 Scalability and Throughput

  • Exact explanations are costly: Shapley attribution scales exponentially, while plausibility- and actionability-constrained counterfactual search can be NP-hard or NP-complete, making approximations important for deployment.Proposed paths include sampling coalitions, independence-assuming Kernel SHAP, and a VAE that generates multiple counterfactuals [165].Source: §5.3.2 · pp. 83–84

5.3.3 Dynamics

  • Most explanation methods assume static black-box models, but distribution shifts are common and only limited work addresses explanations under these changes, [167],.Source: §5.3.3 · p. 84

5.3.4 Security and Privacy

  • With API access, adversaries may combine predictions and explanations to extract approximate models and training-data information via counterfactual augmentation, contrastive boundaries, membership inference, or active interventions yielding causal graphs [169], [134],,.Source: §5.3.4 · pp. 84–85

5.3.5 Robustness vs. Recourse Sensitivity

  • Fokkema et al. show that no attribution method can be both continuous (robust) and recourse-sensitive for some model f, regardless of the utility measure.They identify this as an impossibility result and suggest future work seek workarounds.Source: §5.3.5 · p. 86

Causal Fairness

This section examines causal fairness criteria that assess prediction dependence on protected attributes through causal structure, contrasting counterfactual and interventional approaches and surveying their extensions, applications, and limitations, including whether they rule out unfair disparities and how concrete-data calculations can examine fairness.

Causal Fairness

  • Statistical fairness measures can miss discrimination because they ignore causal relationships, and removing sensitive attributes may fail when other features encode them.Causal graphs can make a feature fair in one setting and unfair in another; statistical measures can also fail under anomalies such as Simpson’s paradox [174].Source: Causal Fairness · p. 87

6.1 Counterfactual Fairness

  • Counterfactual fairness [176] requires an individual’s prediction to remain unchanged in a counterfactual world where only their protected attribute differs.Interventional fairness aims to relax some of the strong assumptions this individual-level criterion requires.Source: Causal Fairness · p. 88, §6.1 · p. 90

6.1.1 Path-Specific Counterfactual Fairness

  • Nabi et al. identify pathway-specific discrimination, while Chiappa requires decisions to match a counterfactual differing in sensitive attributes only along unfair pathways.This finer-grained criterion allows effects through fair pathways, such as gender’s influence on department choice in the Berkeley admissions example.Source: §6.1.1 · pp. 91–92

6.1.2 Causal Explanation Formula

  • The causal explanation formula decomposes decision disparity into direct, indirect, and spurious effects, relating these components to total variation.The stated relationship is TVa0,a1(y) = SEa0,a1(y) + IEa0,a1 (y | a1) −DEa1,a0 (y | a1).Source: §6.1.2 · p. 92, §6.2 · p. 93

6.2 Interventional Fairness

  • Interventional approaches assess population-level effects, including proxy discrimination, and justifiable fairness, which permits influence through admissible variables.K-fairness checks fairness across attribute configurations; proxy fairness intervenes on proxies of protected attributes, and database-repair methods address settings without a known causal graph.Source: §6.2.1 · p. 94, §6.2.2 · p. 94, §6.3 · p. 95

6.3 Fairness under Distribution Shifts

  • Under covariate shift, Singh et al. use causal graphs and conditional independencies to estimate test-set fairness; compound shifts impaired fairness transferability in dermatology and health records [188].Schrouff et al. [188] categorize demographic, covariate, label, and compound shifts, finding clinically plausible shifts can affect multiple distribution aspects simultaneously.Source: §6.3 · pp. 95–96

6.4.1 Alternatives To Equality

  • The agenda extends beyond equal predictive outcomes toward equity, harm, fair recourse, equal effort, and the real-world impacts of decisions, including interference.Proposals include counterfactual harm objectives [193], evaluating effort disparities, and modeling how one person’s outcome depends on decisions about others.Source: §6.4.1 · p. 96, §6.4.2 · pp. 96–97

6.4.3 Partial Identification

  • Causal fairness remains limited by unidentifiable counterfactuals, contested interventions on social categories, conflicting criteria, and models that can introduce unfairness despite fair labels.The authors call for sensitivity analysis under assumption violations; social categories may not admit counterfactual manipulation, criteria can conflict or produce disfavored policies, and introduced unfairness can be studied through path-specific effects.Source: §6.4.3 · p. 97, §6.4.4 · p. 98, §6.4.5 · p. 98

6.4.6 Unfair Models

  • Ashurst et al. find it difficult to rule out unfair disparities even when causal-fairness criteria are met, and the authors call for future research to stress-test those criteria.The paper notes that little work has investigated when fair training labels may still yield an unfair model.Source: §6.4.6 · p. 98
  • Makhlouf et al. [174] provide a more exhaustive example, calculating statistical and causal fairness criteria using concrete data.Source: §6.4.6 · p. 99

Causal Reinforcement Learning … 7.4 Multi-Task RL

This section reviews how causal structure supports decision-making and learning in bandits, model-based reinforcement learning, and multi-task settings. It covers learning from observations and interventions, addressing confounding and spurious state dependencies, and using causal factors or invariant abstractions for exploration and generalization across changing environments.

7.1 Isn’t RL already “Causal”?

  • RL primarily seeks reward-maximizing algorithms, whereas causal inference emphasizes identification and inference; offline RL shares causal inference’s batch-data setting but HTE work prioritizes theoretical guarantees.Offline RL learns policies from trajectories generated by an unobserved policy without online exploration; HTE methods commonly provide convergence rates or confidence intervals.Source: §7.1 · pp. 102–103

7.2 Causal Bandits

  • Causal bandits use observations after interventions and non-interventional data to predict intervention outcomes, potentially improving how quickly agents learn high-reward actions.This framework extends classical and contextual bandits by allowing the learner to use observations beyond the reward and pre-intervention context.Source: §7.2 · p. 104
  • Lattimore et al. show their parallel-bandit and general causal-graph algorithms improve over optimal successive elimination under tested conditions, while leaving general-graph regret bounds open.The parallel-bandit algorithm achieves optimal regret in its setting; lower regret bounds for the more general algorithm remain future work.Source: §7.2 · pp. 105–106
  • Lu et al. prove stronger worst-case regret guarantees for some causal graph types, but meaningful causal gains are impossible for general graphs in the worst case.Bareinboim et al. show that unobserved confounding separates observational from interventional data and propose an agent loss incorporating both distributions.Source: §7.2 · pp. 106–107

7.3.1 Confounded Partial Models

  • Rezende et al. [240] show that partial dynamics models can be confounded by omitted observations and propose backdoor-conditioned Causal Partial Models to address policy-change robustness.The omitted observations influence policy actions, so models that exclude them can plan incorrectly when the behavior policy changes.Source: §7.3.1 · p. 108

7.3.2 Causal World Models

  • Li et al. [241] argue that confounding can make observational and interventional world-model conditionals differ; causal world models target counterfactual outcomes after an abstract-state intervention.The intervention is rendered as a counterfactual observation change, such as object displacement or removal.Source: §7.3.2 · pp. 109–110

7.3.3 Task-Independent State Abstractions

  • Wang et al. introduce causal dynamics learning to reduce spurious dependencies and partition state variables into controllable, action-relevant, and action-irrelevant groups for planning.Action-irrelevant variables, such as a clock in the example, can be omitted; causal predictions avoid unnecessary dependence on other variables such as a second door.Source: §7.3 · p. 111

7.3. Model-Based RL

  • CDL models transition dynamics as a causal graph split by state-variable partitions, using conditional-independence tests and prediction disagreement to learn causal structure.Exploration favors transitions where a dense predictor outperforms the causal predictor, signaling that the learned graph may be inaccurate.Source: §7.3 · pp. 111–112
  • CDL improves generalization and sample efficiency for learned dynamics models and downstream policies.It jointly learns transition dynamics, including the reward function, and plans with that model.Source: §7.3 · p. 112

7.4 Multi-Task RL

  • Multi-task RL studies agents solving multiple environments, including unseen environments with shared invariances or settings that allow limited adaptation.The paper uses task and environment interchangeably.Source: §7.4 · p. 113
  • Causal curiosity learns disentangled representations of causal factors whose effects vary across environments or over time, aiming to make changing behavior interpretable.A factor such as friction can cease to affect action outcomes when a body loses ground contact or when an upward force is applied to a grounded body.Source: §7.4 · p. 113
  • For block MDPs, Zhang et al. learn invariant state abstractions across stochastic observations and single-variable observation interventions, assuming observations identify latent states and state transitions affect only the next state and current reward.The environments share latent states, dynamics, and rewards despite changing observations; an example intervention changes the background color.Source: §7.4 · pp. 113–114

7.4. Multi-Task RL · 7.5 Off-Policy Policy Evaluation · 7.6 Imitation Learning

These sections survey how causal structure informs multi-task reinforcement learning, off-policy policy evaluation, and imitation learning, addressing generalization across environments, distribution shift, confounding, and spurious cues. They describe methods, guarantees, and estimation strategies while identifying conditions and failure modes that limit reliable policy learning.

7.4. Multi-Task RL

  • For an infinite universe of reward-free MDPs, systematic generalization assumes common causal structure and achieves polynomial sample complexity, enabling approximate task solving from finite interactions [254].The task is to provide a slightly sub-optimal policy by planning without additional interactions; variational causal dynamics separately seeks cross-environment adaptation by intervening on learned latent mechanisms.Source: §7.4 · p. 115

7.5 Off-Policy Policy Evaluation

  • Off-policy policy evaluation estimates a target policy’s expected return from behavior-policy trajectories, but distribution mismatch can cause high variance and slow convergence.The text notes that counterfactual inference and OPPE address similar problems and that their techniques may be complementary [216]; causal formulations also enable retrospective evaluation for individual units such as patients.Source: §7.5 · pp. 116–117

7.5.1 Counterfactual Policy Evaluation

  • Counterfactual policy evaluation (CF-PE) can be less biased than model-based evaluation and yields an unbiased estimator of interventional trajectory distributions; it outperformed MB-PE in grid worlds under stated model conditions.The comparison is expected when transition and reward kernels are accurate but the marginal noise distribution is difficult to model; inferred scenarios from observed data can help de-bias the model by excluding implausible noise values.Source: §7.5.1 · pp. 118–120
  • Counterfactually guided policy search represents a POMDP as an SCM and uses counterfactual scenarios inferred from off-policy data to evaluate a target policy.Related work uses counterfactual trajectories to flag episodes with large target-versus-behavior return differences for off-policy debugging in high-risk settings, demonstrating the approach in a sepsis simulation.Source: §7.5.1 · pp. 118–120

7.5.2 Unobserved Confounding

  • Unobserved confounding can distort off-policy policy evaluation, so Kallus et al. minimize worst-case estimated regret over propensity weights representing possible confounding and derive safety guarantees.Their analysis provides uniform control over population regrets consistent with the assumed confounding extent.Source: §7.5.2 · pp. 120–121
  • Off-policy evaluation research also studies partially observable settings and offline-to-online RL, including proximal causal estimators and deconfounding methods for confounded MDPs [266].Other extensions address infinite-horizon evaluation and cases where confounding occurs at only one time step.Source: §7.5.2 · p. 121

7.6 Imitation Learning

  • Causal imitation learning addresses confounding in expert demonstrations by deconfounding observed trajectories or establishing necessary, testable conditions for imitator performance.Imitation learning learns policies from expert examples to reduce environmental interaction or task-specific reward design.Source: §7.6 · pp. 122–123

7.6.1 Causal Confusion

  • Causal misidentification can make behavior cloning perform worse with more data when imitators mistake nuisance variables for causes, as a brake light may be mistaken for the cause of braking [271].A proposed two-stage remedy learns policies for candidate causal graphs, then uses targeted interventions to identify the correct model efficiently.Source: §7.6.1 · pp. 123–124
  • Across three robot tasks, preference-based reward learning remained causally confused despite many pairwise preferences; pedagogic demonstrations improved sample efficiency, but every collection method remained vulnerable [272].Source: §7.6.1 · p. 124

7.6.2 Self-Delusion

  • Conditioning on self-generated actions can cause incorrect causal inferences, whereas treating them as interventions avoids self-delusion; mixing model-generated and external data also yields incorrect updates [273].When the task parameter θ is observed, conditioning and intervening on the action lead to the same prediction over outcomes.Source: §7.6.2 · pp. 124–125, Figure 7.8 · p. 125

7.6.3 Unobserved Confounding

  • With unobserved confounders affecting expert actions and outcomes, imitability is orthogonal to identifiability; Zhang et al. give a graphical criterion, an instrument-checking algorithm, and a policy-estimation procedure.They formalize the setting using partially observable SCMs.Source: §7.6.3 · p. 126

7.6.4 Sequential Data

  • For sequential imitation with multiple decisions per episode, Kumor et al. generalize the backdoor criterion and prove it is necessary for imitability.Source: §7.6.4 · p. 126

7.6.5 Multiple Environments

  • For imitation across environments, Bica et al. seek invariant state features so the learned policy generalizes to new environments.They penalize the energy of the next observation under the policy, encouraging it to stay within high-density regions of the expert’s occupancy measure.Source: §7.6.5 · pp. 126–127, Figure 7.9 · p. 127

7.6.6 Temporally Correlated Noise

  • For temporally correlated confounding in imitation data, Swamy et al. use past states as instruments and propose simulator-based and fully offline IVR estimators.Their performance bounds assume the test-time learner faces the same distribution of temporally correlated noise.Source: §7.6.6 · pp. 127–128

7.7 Credit Assignment … 7.10 Open Problems

This section surveys causal reinforcement-learning approaches to credit assignment, data augmentation, agent incentives, and safety, including methods for attributing rewards, constructing counterfactual experience, and analyzing or constraining agent behavior. It also discusses reward-tampering and delicate-state concerns and identifies open problems in evaluation, formalisms, offline deconfounding, and counterfactual decision-making.

7.7.1 Causal Influence Detection

  • Causal action influence (CAI) uses conditional mutual information to guide exploration and improve sample efficiency and performance when action effects on outcomes are hard to learn.The challenge is acute under partial observability, scale, long horizons, or many actions, where each action may have only a vanishing effect; CAI can also prioritize influential experiences during training.Source: §7.7.1 · pp. 129–130, §7.7 · p. 129

7.7.2 Counterfactual Credit Assignment

  • Counterfactual credit assignment (CCA) conditions value functions on future-event embeddings and uses future-conditional critics to produce unbiased, lower-variance policy-gradient estimates.By comparing action returns while holding many external factors constant, CCA can provide finer-grained, more data-efficient credit assignment; its counterfactuals can be computed model-free.Source: §7.7.2 · pp. 130–131

7.7.3 Multi-Agent Reinforcement Learning

  • In cooperative multi-agent settings, counterfactual causal-influence rewards credit agents for affecting peers’ behavior, encouraging interpretable communication and higher collective rewards than baselines.An agent simulates possible actions and rewards those that significantly affect another agent’s behavior at each time step; some baselines fail to learn completely.Source: §7.7.3 · p. 131

7.8.1 Local Causal Models [290]

  • CoDA augments RL trajectories by swapping factorized subprocesses from trajectories with matching local causal structure, without requiring a forward dynamics model or a particular agent architecture.The method exploits causal independence between subprocesses when they do not interact, as in sparse interactions between billiard balls after the opening break.Source: §7.8.1 · pp. 132–133

7.8.2 Personalized Policies [297]

  • For heterogeneous environments such as healthcare, Lu et al. model transition dynamics as an SCM learned with Bidirectional Conditional GANs and propose algorithms with or without a latent environment variable.When latent heterogeneity is assumed, the causal system explicitly includes an environment variable to represent variation across observations.Source: §7.8.2 · p. 133

7.9.1 Causal Influence Diagrams

  • Causal influence diagrams (CIDs) represent structure, decisions, and utility; with a policy specified, a structural causal influence model supports analysis using interventions and counterfactuals.CIDs can analyze incentives and safety concerns, while value of information measures what a decision-maker would benefit from knowing before deciding.Source: §7.9.1 · pp. 135–136

7.9.2 Incentive Concepts

  • Incentive concepts include materiality, value of information, response incentive, value of control, and instrumental control incentive.They concern, respectively, which observations matter for utility, what information benefits a decision-maker, which variables influence optimal decisions, whether control can increase utility, and whether controlling a variable serves the agent’s goals.Source: §7.9.2 · pp. 136–137

7.9.3 User Interventions

  • Modified-action MDPs model user interventions on an agent’s actions, enabling causal analysis of how RL algorithms adapt to interruptions and whether they try to disable the off-switch.Black-box optimization accounts for environmental causal relationships and may obscure its policy; Q-learning and SARSA with agent-chosen actions ignore modification effects on the current action, while modified-action SARSA considers effects on subsequent actions.Source: §7.9.3 · pp. 137–138

7.9.4 Reward Tampering

  • Everitt et al. use CIDs to analyze reward tampering; proposed link-removal methods avoid undesirable incentives, while approaches differ in whether objectives use observed or latent variables.In the described example, an agent can obtain more rewards by violating the intended task conditions.Source: §7.9.4 · p. 139

7.9.5 Delicate States

  • For delicate states—hard-to-reward, manipulable state components—Farquhar et al. [317] use stable-state conditions and CID criteria to remove dangerous incentives to control them.The stable-state condition is intended to be robust against unmotivated action and make harmful side effects unlikely.Source: §7.9.5 · p. 140

7.10 Open Problems

  • Causal model-based RL methods lack shared evaluation environments, with different studies using distinct benchmarks and procedures.The survey calls for unified evaluation environments to make comparisons possible.Source: §7.10 · p. 140

7.10.1 Unification of Formalisms

  • The authors identify a myriad of MDP formalisms, warn proliferation may cause practitioner decision paralysis and hinder comparison, and recommend rigorous assessment of applicability and possible unification.Examples include MDP with Unobserved Confounders, Confounded MDPs, Causal POMDPs 1 [241], and Causal POMDPs 2, which are not the same formulation.Source: §7.10.1 · pp. 140–141

7.10.2 Deconfounding Offline RL

  • Deconfounding observational data in offline RL is underexplored, although datasets without observed policies or environment access likely contain confounding biases; only a few works address this setup.Source: §7.10.2 · p. 141

7.10.3 Counterfactual Decision-Making

  • Counterfactual inference is neglected in RL, partly because learning correct SCMs is difficult, despite potential benefits for policy-search bias, multi-agent credit assignment, and human-in-the-loop intent.Source: §7.10.3 · pp. 141–142

Modality-specific Applications

This section surveys causal methods for image, text, and graph data, including applications of earlier methods and approaches tailored to specific modalities. It covers uses in vision and NLP such as explanations, augmentation, prediction, robustness, fairness, and model analysis, as well as graph-representation methods addressing invariant prediction, confounding, spurious links, node-specific inference, and connections between graph neural networks and structural causal models.

• Causal Fairness

  • Causal fairness methods aim to make predictions fair with respect to causal relationships involving protected attributes such as gender or race.Source: • Causal Fairness · p. 144

8.1.1 Causal Supervised Learning

  • Tang et al. use backdoor adjustment to remove harmful momentum confounding while preserving the useful mediation path, improving long-tailed image classification and instance segmentation benchmarks.They also compute the intervention p(y | x) to obtain the direct causal effect of X on Y.Source: §8.1.1 · p. 145
  • Yue et al. treat pre-trained knowledge as a confounder in few-shot learning and report that adjustment yields three algorithms with state-of-the-art results on several benchmarks.Their Interventional FSL uses backdoor adjustment to estimate p(y | do(x)); feature-wise, class-wise, and combined adjustments improve baselines across query hardnesses.Source: §8.1.1 · pp. 146, 148
  • Wang et al. propose unsupervised causal attention to adjust for confounders, while Li et al. [333] use a front-door method with an intermediate variable and meta-gradients to improve cross-domain vision performance.The front-door approach avoids identifying confounders when an intermediate variable Z satisfies X → Z → Y.Source: §8.1.1 · pp. 148–150
  • For motion forecasting, Liu et al. [335] separate domain-invariant variables from domain-specific style and content, combining invariant and content representations for predictions intended to generalize across environments.Their architecture uses invariant and style encoders and a style modulator to retain useful domain-specific information while ignoring spurious information.Source: §8.1.1 · p. 150, Figure 8.6 · p. 150
  • Context Adjustment (CONTA) deconfounds pixel-level pseudo-masks in weakly supervised segmentation by simulating p(y | do(x)) with an approximation of unobserved style.It addresses errors from object ambiguity and incomplete backgrounds or foregrounds.Source: §8.1.1 · pp. 150–151
  • Visual grounding adjusts for an unobserved query-location confounder using a learned substitute [337], while video retrieval methods intervene on text/video or query/video inputs to address temporal-location bias,.The video retrieval work identifies that models may exploit moment-location biases rather than cross-modal matching.Source: §8.1.1 · pp. 152–153

8.1.2 Counterfactual Explanations

  • Goyal et al. produce counterfactual visual explanations by identifying how an image must change for a classifier to predict a different class.Source: §8.1.2 · p. 153
  • Counterfactual visual explanations identify image regions or missing attributes that could change a classifier’s decision, and visual counterfactuals can help people distinguish classes.Hendricks et al. [345] generate natural-language explanations of missing evidence; the visual explanations are reported to help human users distinguish classes better.Source: §8.1.2 · pp. 153–154

Counterfactual Data Augmentation

  • Visual counterfactual augmentation intervenes on sample attributes or domains, while V-CDN learns causal interactions and dynamics from video keypoints.Yue et al. [346] vary sample-specific attributes while keeping class attributes fixed and enforce counterfactual faithfulness. Zhang et al. use domain interventions for transferable pose features; Li et al. model interactions beyond collisions, and their stacked-block video work introduces a counterfactual trajectory-prediction benchmark.Source: Counterfactual Data Augmentation · pp. 154–156

8.2.1 Causal Supervised Learning

  • Causal NLP spans explanations, fairness, text classification, and translation; causal VQA trains on counterfactual samples or subtracts the direct language effect from the total effect.Abbasnejad et al. intervene on image or question inputs and report effectiveness on unimodal and multimodal tasks; Niu et al. use the total direct effect rather than the conventional posterior p(a | v, q).Source: §8.2 · p. 156, §8.2.1 · pp. 158, 160

Certified Robustness Against Natural Language Attacks

  • Causal Intervention by Semantic Smoothing (CISS) addresses confounder-driven text vulnerability by smoothing latent semantic content to model p(y | do(x)) and defend against l2-bounded attacks.Zhao et al. [364] motivate the approach by spurious associations between reviewer characteristics and sentiment labels; textual substitutions do not follow Gaussian distributions.Source: Certified Robustness Against Natural… · pp. 160–161

8.2.2 Counterfactual Explanations

  • CausalLM measures concept effects using counterfactual representations, while mediation analysis locates model components causally implicated in gender bias.Garg et al. define counterfactual token fairness and evaluate sensitive-token replacement, counterfactual data augmentation, and counterfactual logit pairing.Source: §8.2.2 · pp. 161–163

8.2.3 Counterfactual Data Augmentation

  • Counterfactual text augmentation improves low-resource entity-recognition generalization and can reduce context-label spurious correlations; machine-translation augmentation preserves context and phrasal alignment.Zeng et al. replace entities with same-type entities and retain examples accepted by a discriminator; Liu et al. use masked and translation language models to replace aligned source and target phrases.Source: §8.2.3 · pp. 163–164

Mitigating Gender Stereotypes in Languages with Rich Morphology

  • For morphologically rich languages, Zmigrod et al. use dependency and morpho-syntactic information to change gendered words and reinflect sentences, reducing gender stereotyping by an average of 2.5 without sacrificing grammar.The unsupervised approach uses dependency trees, lemmata, part-of-speech tags, and morpho-syntactic tags from Universal Dependencies corpora.Source: Mitigating Gender Stereotypes in… · p. 164

8.2.4 Causal Traces

  • Causal tracing [384] finds strong factual-prediction effects in subject-token mid-layer MLP states, and ROME edits targeted MLP weights with a rank-one update.Causal tracing intervenes on hidden states during a forward pass and observes changes in the predicted object; ROME inserts or updates factual associations.Source: §8.2.4 · p. 165, Figure 8.15 · p. 165

8.3.1 Causal Supervised Learning

  • Wu et al. identify invariant rationales by splitting graphs into causal and non-causal subgraphs, intervening on non-causal representations, and training two classifiers with an invariant risk loss.The method is proposed to address spurious correlations that can undermine intrinsic interpretability and out-of-distribution performance.Source: §8.3.1 · p. 167
  • Chen et al. [393]’s GOOD improves out-of-domain generalization by separating invariant subgraphs from shift-sensitive parts and aligning causal mechanisms through a featurizer and classifier.The featurizer extracts an invariant subgraph, which the classifier uses to predict the label.Source: §8.3.1 · pp. 167–168
  • Sui et al. use attention modules to separate critical from trivial graph parts and apply backdoor adjustment to mitigate confounding by trivial subgraphs.Source: §8.3.1 · p. 168
  • Zhao et al. [396] augment link-prediction training with counterfactual links to reduce reliance on spurious graph-structure associations.Their counterfactual query asks whether a link would still exist if the graph structure were different from what was observed.Source: §8.3.1 · p. 169

8.3.3 Miscellaenous

  • Feng et al. assess neighborhood effects by comparing original and no-neighbor predictions, using causal-effect uncertainty and a threshold to decide whether to retain a node in training data.High uncertainty can indicate that a test node’s local structure differs from observed properties and its predicted class may be incorrect.Source: §8.3.3 · pp. 169–171
  • Zečević et al. [399] formalize interventions for graph neural networks in a neural-causal model class and analyze its feasibility, expressivity, and identifiability for causal inference.Source: §8.3.3 · p. 171

Causal Benchmarks

CausalML benchmarks provide interventional or counterfactual ground truth across reinforcement learning, vision, and language, while conventional benchmarks can sometimes be repurposed but are not suitable for every causal task.

Causal Benchmarks

  • Conventional OOD and RL benchmarks can sometimes evaluate causal methods, but causal comprehension requires purpose-built question-answer pairs, and NLP benchmark construction is hindered by large vocabularies and scarce intervention ground truth.Conventional RL simulations can be interpreted as SCMs, whereas not all conventional ML benchmarks suit the techniques discussed.Source: Causal Benchmarks · p. 172, §9.3 · p. 177

9.1. Reinforcement Learning

  • CausalRL benchmarks range from CausalWorld’s intervention API and shared causal task structures [10] to parameterizable discovery environments, procedurally resampled Alchemy, and CausalCity’s driving scenarios.CausalWorld lets users or agents intervene on environment variables; Alchemy requires identifying relevant latent structure, while CausalCity models high-level vehicle agency and complex causal scenarios including confounders.Source: §9.1 · pp. 173–174

9.2 Computer Vision

  • CLEVRER tests descriptive, explanatory, predictive, and counterfactual video reasoning, and state-of-the-art models perform poorly on its causal tasks.The authors argue that future methods must learn the underlying causal relations between objects in images.Source: §9.2 · p. 175
  • CoPhy provides ground-truth counterfactual trajectories after interventions on initial conditions, while Filtered-CoPhy [351] tests counterfactual trajectory prediction across three video types.CoPhy covers physical dynamics such as falling block towers, bouncing balls, and collisions; Filtered-CoPhy includes BlocktowerCF, BallsCF, and CollisionCF, with confounders designed to be recoverable by sufficiently powerful representations.Source: §9.2 · p. 176
  • 70%: robustness methods cannot achieve over 70% accuracy on Spawrious’s hardest split with an ImageNet-pretrained ResNet50 [36].Causal3DIdent supports evaluation of style and content decomposition methods by providing images of 3D objects with causal dependencies.Source: §9.2 · p. 176

9.3. Natural Language Processing

  • Counterfactually-Augmented Data, CausaLM, and CRASS [409] provide NLP testbeds for counterfactual label changes, concept removal, and counterfactual reasoning, respectively.Counterfactually augmented documents are human-revised to match target labels while preserving unrelated facts; CausaLM includes ground-truth counterfactual examples, and CRASS is part of BIG-bench.Source: §9.3 · p. 177
  • 74.1% vs. 90.53% accuracy: fine-grained causal reasoning models trail the human ceiling, indicating persistent difficulty with causal reasoning [411].The dataset covers causality detection, fine-grained causality extraction, and causal QA; finer-grained events such as enabling and preventing effects matter for answering what-if questions.Source: §9.3 · p. 178

The Good, the Bad and the Ugly

The section weighs CausalML’s benefits across learning, generation, explanations, fairness, and reinforcement learning against practical challenges and risks from scarce evaluation data and untestable assumptions.

10.1 The Good

  • Causal supervised learning targets domain-robust, reusable features or mechanisms [11], while causal generative methods enable controllable counterfactual generation using causal graphs and weaker domain knowledge.Source: §10.1 · pp. 179–180
  • Causal explanations provide feature attributions or actionable counterfactuals, while causal fairness uses interventional or counterfactual criteria to assess models and address selection bias.Source: §10.1 · p. 180
  • Causal reinforcement learning uses interventions to formalize actions and environmental changes, deconfound trajectories, model reward effects, and support counterfactual data augmentation.Source: §10.1 · p. 180

10.2.1 Lack of Open-Source Ecosystem

  • CausalML lacks the convenient SCM APIs, identification tests, manipulation tools, benchmarks, and pretrained SCMs available in the broader open-source ML ecosystem.The authors note that ingredients for CausalML platforms are ready, including algorithms that automatically verify identifiability.Source: §10.2.1 · p. 181

10.2.2 Lack of Comparisons to Non-Causal Methods

  • CausalML studies often omit non-causal baselines on the same metrics, weakening progress assessment and leaving practitioners without fair comparisons.Source: §10.2.2 · p. 182, §10.3 · p. 183

10.3.1 Difficulties of Obtaining Ground-Truth Evaluation Data

  • Ground-truth causal benchmarks are scarce because real-world interventions can be costly, unethical, or time-consuming, and real-world counterfactuals are unobservable.Tailored, inconsistent simulations hinder comparison, although the authors suggest that more RCT datasets and simulators may help as the field matures.Source: §10.3.1 · pp. 183–184

10.3.2 Untestable Assumptions Are Inevitable

  • SCM assumptions can amplify bias and harm external validity; adjusting for covariates acting like instruments may increase residual confounding bias, so simpler models can work better.Source: §10.3.2 · p. 184, §10.3 · p. 185

Related Work

The section first reviews surveys applying causal reasoning across machine learning domains, then turns to machine learning methods for causal inference, especially causal-effect estimation and causal discovery.

11.1 Other Surveys

  • Related surveys connect causal inference to machine learning robustness and modularity, NLP generalization, fairness and interpretability, socially responsible AI, and visual representation learning [11], [445], [448], [449].Liu et al. [449] categorize visual work into causal comprehension, robustness, question answering, and datasets, and identify future directions including better causal modeling, intervention approximations, counterfactual synthesis, and benchmarks.Source: §11.1 · pp. 186–187

11.2 Machine Learning for Causal Inference

  • The survey distinguishes using causal theory to address machine-learning problems from using machine learning to answer causal questions, including effect estimation and discovery.It briefly reviews advances in those two causal-inference tasks for completeness.Source: §11.2 · pp. 187–188

11.2.1 Causal Effect Estimation

  • Effect estimation must handle missing interventions and confounding; with observed confounders, treatment imbalance can bias ordinary regression, motivating regularization and meta-learners,,,.Neural meta-learners address binary, continuous, higher-dimensional, and arbitrary treatments, with approaches including splines, propensity features, and Transformers,,, [460].Source: §11.2.1 · pp. 188–189
  • For unobserved confounding, effect-estimation strategies include partial-identification bounds, sensitivity analysis, and instrumental or proxy variables; proxy-based methods seek less biased estimates than direct adjustment,,,,.A proxy may encode auxiliary information about a hidden confounder, such as reservation-page views serving as a proxy for desire to fly.Source: §11.2.1 · pp. 189–190
  • Instrumental-variable methods require an instrument independent of the confounder, associated with treatment, and conditionally independent of outcome; one method also addresses treatment measurement error given an instrument,.The passage gives oil prices as a possible instrument for ticket prices, affecting sales through price.Source: §11.2.1 · p. 190

11.2.2 Causal Discovery

  • Causal discovery seeks structures from observational or interventional data, but faces non-identifiability and super-exponential search spaces; methods include combinatorial, continuous-relaxation, and permutation-based approaches,,,,.Outputs can range from partial node orderings to full SCMs under assumptions such as linear structural equations.Source: §11.2.2 · pp. 190–191
  • To avoid full-graph recovery in high dimensions, work targets ancestral structures, partial ancestral graphs, or selected causal orders, while active discovery seeks data efficiency and fewer interventions,,,.Granger causality quantifies whether a time series' past helps predict another's future; recovering a DAG requires observing relevant variables and having no instantaneous effects, while nonlinear extensions exist, [7], [513].Source: §11.2.2 · p. 191, §11.2.3 · pp. 191–192

Conclusion

The conclusion presents causal inference as a basis for interventional and counterfactual reasoning, summarizes its applications across ML areas and modalities, and identifies open problems and existing benchmarks.

Causal Inference

  • Causal inference encodes assumptions about a system, enabling reasoning about interventional and counterfactual estimands beyond conventional statistical or probabilistic inference.Source: Conclusion · p. 194

Causal Supervised Learning

  • Causal supervised learning improves predictive generalization by learning invariant features or mechanisms that reduce reliance on spurious associations.Future work includes targeted invariance benchmarks, links to adversarial robustness and meta-learning, and use of additional supervision signals.Source: Conclusion · p. 194

Causal Generative Modeling

  • Causal generative modeling samples from interventional or counterfactual distributions for controllable generation or sample editing.Methods learn structural assignments and some infer causal structure; open issues include abstraction levels, scaling assignment learning to larger graphs, and when counterfactual data augmentation works.Source: Conclusion · p. 194

Causal Explanations

  • Causal explanations account for the causal structure of model mechanics or data generation, using feature attributions or contrastive explanations.Open problems include unifying and scaling methods, robustness to distribution shifts, security and privacy, and the trade-off between robustness and recourse sensitivity.Source: Conclusion · p. 195

Causal Fairness

  • Causal fairness assesses and mitigates disparities using criteria grounded in causal relationships in the underlying data.The criteria rely on counterfactual or interventional distributions; future work includes alternatives to equality, fairness beyond standard prediction, weaker observability assumptions, and the validity of interventionist views on social categories.Source: Conclusion · p. 195

Causal Reinforcement Learning

  • Causal reinforcement learning methods account for the decision environment’s causal structure, with claimed benefits including deconfounding, intrinsic rewards, and data efficiency.The conclusion identifies possible unification of formalisms, largely unaddressed deconfounding of offline data, and potential benefits from counterfactual decision-making agents.Source: Conclusion · p. 195

Modality Applications

  • Causal and modality-specific principles offer opportunities to improve computer vision, natural language processing, and graph representation learning.Source: Conclusion · p. 195

Causal Benchmarks

  • The survey reviews CausalML benchmarks, including benchmarks with ground-truth causal interventions and counterfactuals.Source: Conclusion · p. 195

192 Conclusion · Appendices

The conclusion reviews CausalML’s benefits and identifies methodological and practical challenges, then briefly situates the survey among two categories of related work.

192 Conclusion

Related papers

Loading 2206.15475v3…