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

Causal machine learning faces challenges in generalization, causal identifiability, explanations, and evaluation resources. This paper surveys methods across major CausalML problems and modalities, finding both opportunities for improved predictive generalization and unresolved questions about scalable, robust causal explanations.

  • Problem

    CausalML faces limited robustness to distribution shifts, hidden confounding, and scarce public benchmark resources for evaluating causal models.

  • Method

    The paper systematically categorizes and compares CausalML methods, reviews modality-specific applications, and discusses benchmarks and open problems.

  • Results

    The survey reports improved predictive generalization from invariant-feature and invariant-mechanism approaches, while causal explanations remain difficult to unify, scale, and robustly deploy.

  • Takeaways & Limitations

    CausalML offers modality-specific opportunities for computer vision, natural language processing, and graph representation learning, alongside substantial open problems.

  • Takeaways & Limitations

    Many reviewed methods rely on assumptions that may fail in practice or cannot be tested when data lack sufficient interventions, variety, or domain knowledge about unobserved confounders.

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.

Causal Machine Learning

Causal Machine Learning (CausalML) is presented as an umbrella term.

  • Causal Machine Learning (CausalML) is described as an umbrella.
  • The passage introduces CausalML as a broad umbrella concept.
  • CausalML is framed as an umbrella rather than a narrowly defined method.

Introduction

The survey frames CausalML as a response to real-world ML challenges that causal formalisms can help address, using Structural Causal Models to reason about interventions and counterfactuals. It organizes existing work into problem classes, reviews modality-specific applications and benchmarks, and identifies future research directions.

  • Motivation: Real-world deployment exposes distribution-shifted generalization, limited generative control, unfair predictions, abstract interpretability, and unstable reinforcement-learning transfer.These challenges motivate causal approaches to open ML problems.
  • Causal framework: Structural Causal Models formalize the data-generating process and support reasoning about interventions and counterfactual consequences.Interventions describe changes to the generating process, while counterfactuals model consequences in hindsight while accounting for what happened.
  • Motivation: The survey addresses the lack of clear CausalML categorization, whose assumptions can be unfamiliar, difficult to relate across setups, and often untestable.These difficulties make progress and applicability hard to measure.
  • Survey scope: It provides a self-contained introduction to causality and taxonomizes existing work into causal supervised learning, causal generative modeling, causal fairness, and additional problem classes.The survey compares methods for each problem class and discusses future-work avenues.
  • Survey scope: It reviews applications in computer vision, natural language processing, and graph representation learning, alongside causal benchmarks.These reviews are data-modality-specific and form part of the survey’s broader coverage.

4 Introduction

The paper evaluates CausalML’s benefits over non-causal machine learning, its current challenges, and the costs of using causal techniques.

  • 4 Introduction: The paper discusses CausalML’s benefits compared with non-causal ML, the field’s challenges, and the inevitable costs of applying causal techniques.These themes are addressed as “the good,” “the bad,” and “the ugly,” respectively.

Causality: A Minimal Introduction

Causal reasoning is formalized with causal DAGs, which encode structural assumptions and support reasoning about interventions and counterfactuals. DAGs also express conditional independencies that simplify joint-distribution modeling, while hidden confounding can cause machine-learning models to rely on spurious associations.

  • Causal DAGs: Causal DAGs encode a priori assumptions about causal structure and enable reasoning about interventions and counterfactuals.They provide a canonical representation of causal relations.
  • Causal DAGs: A DAG is a directed graph without directed cycles, with edges from parents to children and directed paths defining ancestors and descendants.Parents are denoted pa(X), ancestors an(Y), and descendants de(X).
  • Graphical Models: DAG-based factorization replaces naive joint modeling with smaller conditional factors, reducing dependence on all preceding variables.For the farmer example, the joint distribution is simplified to p(t)p(d | t)p(r | t)p(s | d, r).
  • Graphical Models: The Markov condition makes each node independent of its non-descendants given its parents, yielding factorization into parent-conditioned distributions.This decomposes complex joint functions into smaller conditional factors according to the assumed DAG relations.
  • Hidden Confounding: Hidden confounding can lead sophisticated machine-learning models to rely on spurious training associations that may disappear in production.The text notes that this reliance is a feature unless the model is explicitly constrained to avoid such associations.

Causal Supervised Learning

Causal supervised learning addresses failures of the i.i.d. assumption by modeling data across environments as interventional distributions and seeking predictive features that remain invariant. Its claims depend on strong, often unverifiable SCM and decomposition assumptions, motivating robustness benchmarks and further work on limited interventions and unobserved confounding.

  • Causal formulation: Causal supervised learning instead treats environments as samples from separate interventional distributions governed by a structural causal model.This reframes distribution shifts, such as interventions on latent background variables, within an SCM-based perspective.
  • Motivation: i.i.d. training can exploit spurious feature-label correlations, causing poor performance when test distributions contain changed frequencies or novel feature combinations.Examples include cows photographed against alpine backgrounds and one-to-one or many-to-many correlations between features and classes.
  • Causal representations: Causal representation methods simplify SCMs or assess whether representations capture non-spurious and efficient features, while contrastive augmentations can simulate style interventions under style-content assumptions.These approaches rely on assumptions such as no hidden confounding among the grouped variables.
  • Invariant prediction: Invariant Risk Minimization learns features that are predictively invariant across environments and combines them with a linear classifier that is optimal in every environment.IRM extends invariant causal prediction by learning a feature map ϕ(X) and classifier w whose composition predicts Y.
  • Benchmarks: Synthetic photo-realistic benchmarks with controlled spurious correlations and explicit distribution shifts enable systematic evaluation of robustness methods.These datasets can include multiple difficulty levels while preserving visual complexity through modern text-to-image models.
  • Open problems: The field faces a gap between idealized causal assumptions and what finite data can establish, especially when interventions are limited or unobserved confounders obscure genuine causal paths.Additional domain knowledge may be needed to distinguish correlations from causal relationships, and assumptions may be difficult or impossible to verify observationally.

Causal Generative Modeling

Causal generative modeling enables controllable generation, sample editing, and counterfactual data augmentation by modeling interventional or counterfactual distributions. Its methods depend on supervision, domain knowledge, and choices about causal structure and abstraction, leaving scalability, representation learning, and benchmarking as open problems.

  • Causal Generative Modeling: Causal generative models replace observational conditioning with interventional or counterfactual distributions for controllable generation and sample editing.Given causal representation Z, controllable generation samples from p(x | do(a)), while editing targets p(x_A→a | x).
  • Applications: Counterfactual generation supports medical-imaging mechanism studies and augmentation that changes spurious features, such as cow backgrounds, to improve classifier generalization.DeepSCM generates brain MRI counterfactuals under changed biological sex; Sauer et al. generate cows in unfamiliar backgrounds, and counterfactual augmentation improves classifier performance more than non-counterfactual GAN samples.
  • Assumptions and Limitations: Learning causal generative models generally requires domain knowledge or supervision because disentangled representations cannot be learned unsupervisedly without extensive domain knowledge.Common assumptions include a known or well-approximated graph, no hidden confounders, correct causal ordering, and an appropriate abstraction level.
  • Causal Representation Learning: Causal representation learning is ill-posed without specifying the SCM’s abstraction level, and the efficient, interpretable choice is difficult to determine automatically.Pixel-level and object-level SCMs can perform equally with sufficient data and compute, while existing approaches require specifications such as variable labels, latent dimensionality, or intervention pairs.
  • Open Problems: Scaling causal generative models beyond graphs of order 5 and evaluating whether counterfactual sampling improves external predictors remain open problems.DCEVAE is the exception among reviewed structural-assignment approaches, while VACA faces GNN over-smoothing; CFDA benchmarking is under-explored.

Causal Explanations

The section focuses on model-agnostic, local causal explanations through feature attribution and contrastive explanations. It highlights methods that incorporate causal structure to improve attribution or generate feasible, causally effective recourse actions.

  • Causal Explanations: Causal explanations here comprise local, model-agnostic feature attribution and contrastive techniques, excluding global and model-specific explanations.They explain individual predictions from a single input; global methods describe dataset-wide feature reliance, while model-specific methods include neural-network saliency maps.
  • Feature Attribution: Causal feature attribution distinguishes features that causally affect predictions from merely correlated features, directing applicants toward more effective changes.In the loan example, causal attribution would prioritize improving income rather than changing correlated features such as education, job, or industry.
  • Feature Attribution: Asymmetric Shapley values incorporate incomplete causal knowledge while relaxing Shapley symmetry, spanning data-agnostic Shapley values and causality-based explanations.They permit known causal information to influence explanations without requiring full causal inference.
  • Feature Attribution: Causal Shapley values address a conceptual flaw in standard Shapley explanations by explaining the total effect of features.The supplied passage introduces CSVs as a remedy but does not provide further formulation details.
  • Contrastive Explanations: Causal counterfactual explanations must account for feasibility and action effects, because nearest counterfactuals or decision-boundary optimization can yield suboptimal recourse.Plausibility constraints can encode feasibility or diversity, while consequential recommendations exploit causal effects of actions on predictions.

Causal Fairness

Causal fairness uses causal structure to assess and mitigate discrimination that statistical criteria can miss, including when sensitive information is encoded redundantly in other features. The section covers counterfactual and interventional criteria, path-specific refinements, and trade-offs among competing fairness requirements.

  • Motivation: Causal structure is essential because statistical fairness measures can miss discrimination under different causal relationships, including Simpson’s paradox.Whether an input is fair to include depends on the causal graph generating the data.
  • Motivation: Removing or down-weighting sensitive attributes may remain unfair because other features can redundantly encode them, especially in large feature spaces.A classifier can learn an implicit representation of sensitive features even when those attributes are not explicitly provided.
  • Fairness criteria: Causal fairness criteria comprise counterfactual fairness, which evaluates protected-attribute changes in counterfactual worlds, and interventional fairness, which relaxes strong counterfactual assumptions.A counterfactually fair predictor is invariant when only the protected attribute changes and all else remains equal.
  • Path-specific fairness: Path-specific counterfactual fairness refines whole-attribute invariance by requiring decisions to remain unchanged when sensitive attributes differ only along unfair causal pathways.This distinguishes potentially fair effects, such as gender operating through department choice, from unfair pathways.
  • Limitations: Fairness criteria can conflict, so multiple admissible requirements may be impossible to satisfy simultaneously and must be balanced.Prior work explores balancing certain measures with a multi-objective framework while also identifying limitations and adverse consequences of causal fairness.

Causal Reinforcement Learning … 7.4 Multi-Task RL

Causal Reinforcement Learning applies causal reasoning to reinforcement-learning problems, addressing decision-making from observational data, confounding, model robustness, and generalization across environments. The reviewed methods exploit causal structure to improve learning, planning, state abstraction, and transfer.

  • Causal Reinforcement Learning: Causal Reinforcement Learning uses causal methods to benefit reinforcement-learning problems, whose agents learn actions by maximizing rewards through interaction with environments.RL is commonly formalized with Markov decision processes containing observations, actions, and rewards.
  • 7.1 Isn’t RL already “Causal”?: RL primarily maximizes expected rewards, whereas causal inference emphasizes identifying and estimating effects from limited observational data; offline RL lies between these settings.Offline RL learns policies from trajectories generated by unknown policies without online exploration, resembling causal inference’s data constraints.
  • 7.2 Causal Bandits: Causal bandits use observational information and causal structure to predict intervention outcomes, potentially improving the rate at which policies learn high-reward actions.This framework generalizes classical bandits by exploiting non-interventional data from a fixed causal model.
  • 7.2 Causal Bandits: Causal-bandit algorithms achieve optimal regret in parallel bandits and improve over optimal successive elimination by exploiting causal knowledge, while general-graph lower regret bounds remain open.For some graph classes, causal algorithms have stronger worst-case regret guarantees than non-causal methods, but general causal graphs may offer no worst-case advantage.
  • 7.2 Causal Bandits: Unobserved confounders make observational and interventional data distinct, so maximizing rewards under the observational distribution is not always optimal.Causal formulations therefore modify the conventional multi-armed-bandit objective to account for this distinction.
  • 7.3.1 Confounded Partial Models: Partial model-based RL dynamics can be causally incorrect because omitted observations confound policy-generated actions, making planning non-robust to behavior-policy changes.Causal Partial Models use backdoor adjustment, whereas Non-Causal Partial Models do not condition on the relevant backdoor variables.
  • 7.3. Model-Based RL: Causal Dynamics Learning learns action and state dependencies for task-independent state abstraction, producing predictions less reliant on spurious associations and improving generalization and sample efficiency.The method partitions variables into controllable, action-relevant, and action-irrelevant groups and jointly learns dynamics and planning policies.
  • 7.4 Multi-Task RL: Multi-Task RL studies agents solving multiple environments, including settings with unseen environments, shared invariances, or limited adaptation phases.Causal approaches learn disentangled causal factors, invariant state abstractions, sparse and invariant representations, and reusable cause-effect schemas for systematic generalization.

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

These sections develop causal approaches to generalizing across RL environments, evaluating policies from off-policy data, and learning imitation policies robust to confounding, causal misidentification, and distribution shifts. Across them, SCMs, counterfactual inference, causal structure, and instrumental variables provide formal tools for efficient generalization, safer evaluation, and more reliable imitation.

  • 7.4. Multi-Task RL: Multi-task RL formalizes systematic generalization as solving any reward-specified task over an environment universe after finitely many interactions, without further interactions.The framework assumes common causal structure across transition dynamics, enabling a provably efficient algorithm with polynomial sample complexity.
  • 7.4. Multi-Task RL: Variational Causal Dynamics learns sparse, causally factorized latent dynamics so novel environments can be modeled by intervening on independent latent mechanisms.The method jointly learns causal adjacency and environment-specific intervention masks using variational inference and an ELBO objective.
  • 7.5. Off-Policy Policy Evaluation: Off-policy policy evaluation estimates a target policy’s expected return from behavior-policy trajectories, but distributional mismatch makes the task difficult.Its appeal includes reusing existing experience when online experimentation is costly or ethically constrained and improving data efficiency.
  • 7.5.1 Counterfactual Policy Evaluation: Counterfactual policy evaluation represents POMDPs as SCMs and uses posterior-inferred scenarios to reduce bias relative to model-based evaluation when transition models are accurate but noise distributions are difficult to model.Counterfactual inference yields an unbiased estimator of the interventional trajectory distribution, and CF-PE outperforms MB-PE in partially observed grid-world settings.
  • 7.5.2 Unobserved Confounding: Unobserved confounding can undermine OPPE, motivating robust policy improvement methods that minimize worst-case regret and provide safety and uniform regret-control guarantees.Related work extends these ideas to infinite-horizon and partially observable settings, combines offline and online data, and connects OPPE to instrumental-variable regression.
  • 7.6. Imitation Learning: Imitation learning learns control policies from expert demonstrations to reduce environmental interaction and task-specific reward design, but confounding can prevent recovery of expert performance.Causal imitation learning therefore seeks to deconfound observed trajectories or establish testable conditions for matching expert behavior.
  • 7.6.1 Causal Confusion; 7.6.2 Self-Delusion: Causal confusion causes cloned policies to misidentify nuisance variables as causes, so more observations can worsen performance; interventions and causal-model search are proposed remedies.In the driving example, a model relying on a brake indicator performs poorly when that effect is unavailable, despite low training loss.
  • 7.6.2 Self-Delusion; 7.6.3 Unobserved Confounding; 7.6.4 Sequential Data; 7.6.5 Multiple Environments; 7.6.6 Temporally Correlated Noise: Other causal imitation settings address self-delusion, unobserved confounding, sequential and multi-environment imitation, and temporally correlated noise using interventions, graphical criteria, invariant features, energy objectives, and instrumental variables.Treating self-generated actions as interventions prevents backward information flow, while past states can serve as instruments for deconfounding state-action associations.

7.7 Credit Assignment … 7.10 Open Problems

The paper surveys causal approaches to reinforcement learning, covering credit assignment, counterfactual data augmentation, and incentive analysis, then identifies open problems involving evaluation, formalism unification, offline deconfounding, and counterfactual decision-making. Across these topics, causal models clarify action effects, support counterfactual reasoning, and expose safety-related incentives.

  • 7.7 Credit Assignment: Credit assignment links actions to rewards while separating an agent’s skill from luck, but partial observability, scale, long horizons, and many actions can make effects vanishingly small.Causal influence and counterfactual methods address this difficulty by estimating action-specific causal effects.
  • 7.7 Credit Assignment: Causal action influence improves exploration and training by prioritizing states and experiences with stronger predicted influence, while counterfactual credit assignment yields finer-grained, more data-efficient estimates.CCA conditions value functions on future event embeddings and estimates returns while holding external factors constant; its counterfactuals can be computed model-free.
  • 7.7 Credit Assignment: In cooperative multi-agent reinforcement learning, causal influence rewards use counterfactual reasoning to credit agents for effects on other agents in addition to shared rewards.Shared global rewards otherwise make each agent’s contribution difficult to infer.
  • 7.8 Counterfactual Data Augmentation: Counterfactual data augmentation exploits locally independent causal mechanisms to swap factorized trajectory subspaces without requiring a forward dynamics model.The approach infers local interactions and augments data when trajectories share the same local factorization.
  • 7.9 Agent Incentives: Causal influence diagrams and structural causal influence models provide graphical formalisms for analyzing agent incentives, interventions, utility, information, and counterfactuals.A policy turns a structural causal influence model into an SCM, enabling standard causal interventions and counterfactual definitions.
  • 7.9 Agent Incentives: Causal incentive analysis addresses safety problems including user interventions, reward tampering, and control of delicate states.Modified-action MDPs expose how interruptions affect causal assumptions, while reward-tampering analyses identify incentives to manipulate implemented rewards.
  • 7.10 Open Problems: Open problems include non-overlapping evaluation environments, redundant formalisms, under-explored deconfounding in offline reinforcement learning, and neglected counterfactual inference.The paper recommends more rigorous comparisons and possible unification of formalisms, alongside greater attention to counterfactual decision-making.
  • 7.10 Open Problems: Counterfactual reasoning may reduce policy-search bias, assign credit in shared-reward systems, and account for intended human actions in human-in-the-loop settings.The paper also highlights causal modeling for large-scale systems such as ad placement as a motivation for carefully designed counterfactual agents.

Modality-specific Applications

This section reviews causal machine learning methods for image, text, and graph data, including both applications of earlier causal methodologies and modality-specific approaches. Across these modalities, methods address spurious associations, confounding, counterfactual reasoning, invariant features, and fairness.

  • Overview: The chapter organizes modality-specific methods around image, text, and graph data, combining applications of earlier causal methods with approaches tailored to each modality.The reviewed modalities are computer vision, natural language processing, and graph representation learning.
  • Computer Vision: Vision methods use causal adjustment, interventions, and confounder control to improve long-tailed recognition, few-shot learning, cross-domain generalization, and weakly supervised segmentation.Examples include backdoor adjustment for long-tailed classification and segmentation, adjustment for pre-trained knowledge in few-shot learning, interventions on intermediate variables, and CONTA for pseudo-mask deconfounding.
  • Natural Language Processing: Text applications examine counterfactual fairness, factual model editing, and causal biases introduced by dataset collection and sentence-pair direction.Counterfactual fairness asks how classifier outputs change when sensitive tokens differ, while ROME inserts or updates factual associations through rank-one weight updates.
  • Graph Representation Learning: Graph methods learn invariant subgraphs, deconfound critical and trivial graph components, and train on counterfactual structures to reduce spurious correlations and improve out-of-domain performance.These approaches target biases in intrinsic interpretability, domain shifts, graph-level confounding, and link-prediction assumptions about graph structure.

Causal Benchmarks

CausalML benchmarks provide interventional or counterfactual ground truth across reinforcement learning, vision, and language, while exposing persistent gaps in causal reasoning and limitations of conventional benchmarks. They include controllable simulators, causal datasets, and counterfactual evaluation tasks designed to test whether models learn underlying causal relationships.

  • Benchmark scope: CausalML benchmarks are designed around interventional or counterfactual ground-truth data, but conventional benchmarks remain useful when their tasks can be interpreted causally.Standard OOD benchmarks and RL simulators such as MuJoCo and the Arcade Learning Environment can evaluate some causal methods, whereas causal comprehension requires explicitly causal question-answer pairs.
  • Causal reinforcement learning benchmarks: CausalWorld, Visual Causal Discovery environments, Alchemy, and CausalCity provide controllable settings for interventions, latent causal-variable discovery, procedurally varying structures, and complex driving scenarios.CausalWorld exposes an API and causal graph for environment interventions; Alchemy resamples latent structures each episode; CausalCity models high-level agency and confounding in driving.
  • Computer vision benchmarks: Vision benchmarks cover causal video reasoning, physical dynamics, counterfactual trajectory prediction, causal disentanglement, and robustness to spurious correlations.CLEVRER exposes poor performance by state-of-the-art visual reasoning models, CoPhy tests synthetic physical reasoning, Filtered-CoPhy tests counterfactual trajectories, and Causal3DIdent studies style-content decomposition.
  • Computer vision benchmarks: Spawrious shows that current robustness methods remain below 70% accuracy on its hardest split using an ImageNet-pretrained ResNet50.The benchmark suite introduces spurious correlations between image classes and backgrounds.
  • Natural language processing benchmarks: NLP benchmarks evaluate counterfactual rewriting, concept removal, counterfactual question answering, and fine-grained causal reasoning, with models reaching 74.1% versus a 90.53% human ceiling.The datasets span sentiment analysis, natural language inference, causal explanations, counterfactual reasoning, causality detection, causal extraction, and Causal QA.

The Good, the Bad and the Ugly

CausalML offers formal tools for robust prediction, controllable generation, explanations, fairness, and reinforcement learning, but its progress is constrained by limited software, weak non-causal comparisons, and scarce public benchmarks. More complex causal models can also introduce new bias or practical costs.

  • The Good: Causal supervised learning supports domain-robust, reusable features and mechanisms by modeling invariances and intervention-induced distribution changes.Invariant feature learning targets causal parents of Y, while invariant mechanism learning models changes associated with independent unobserved confounders.
  • The Good: Causal generative modeling enables controllable generation by incorporating domain knowledge and producing counterfactual samples that respect causal dependencies.Structural assignment learning adds knowledge of the underlying causal graph for the data-generating process.
  • The Good: Causal explanations identify intervention-relevant variables or generate contrastive predictions based on actionable counterfactual inputs, while causal fairness evaluates and enforces criteria involving sensitive attributes.Fairness criteria can deconfound potential sources of selection bias when causal dependencies are available.
  • The Good: Causal reinforcement learning uses interventions to formalize actions, environmental changes, trajectory deconfounding, reward effects, state-space changes, and counterfactual data augmentation.These uses can expose opportunities to recycle already observed information, although the supplied passage truncates the final claim.
  • The Bad: CausalML lacks a comparably convenient open-source ecosystem for SCM identification, manipulation, benchmarks, and pretrained models, despite available packages for selected causal estimation tasks.The paper argues that APIs for automatically verifying identifiability would be especially helpful.
  • The Bad: Many CausalML studies omit comparisons with non-causal methods on shared metrics, while public interventional benchmarks remain scarce because real-world interventions require difficult active experimentation.Reported omissions include non-causal out-of-distribution, multitask reinforcement-learning, and other baseline comparisons.
  • The Ugly: More elaborate causal adjustment can amplify residual confounding bias, so a simpler model excluding treatment-predictive covariates may perform better.Pearl’s example concerns covariates acting like instrumental variables under hidden confounding.

Related Work

Related work connects causal inference with machine learning across representation learning, NLP, visual learning, healthcare, and socially responsible AI. It also reviews how modern machine learning estimates causal effects and addresses confounding, identification, and robustness challenges.

  • Foundational work links machine learning with graphical causal inference and emphasizes distribution-shift robustness and reusable, modular mechanisms.
  • Causal formalisms are proposed to make NLP methods more robust and understandable by addressing spurious associations, group performance differences, and inscrutability.
  • Visual causal representation learning is organized into comprehension, robustness, visual question answering, and causal datasets, with future needs for better modeling, interventions, counterfactuals, and benchmarks.
  • Healthcare applications use causal machine learning for clinical decision support and medical image analysis, targeting high-dimensional data, distribution shifts, temporal relationships, and failures to distinguish correlation from causation.
  • Modern representation learning supports causal effect estimation and discovery, while effect-estimation methods address missing interventions, confounding, regularization, meta-learning, partial identification, sensitivity analysis, proxies, and instruments.

Conclusion

The conclusion argues that causal assumptions enable reasoning about interventional and counterfactual estimands, benefiting several machine-learning areas. It highlights advances and open problems across causal learning, modality-specific applications, and benchmarking.

  • Conclusion: Causal inference encodes system-level causal assumptions, enabling reasoning about interventional and counterfactual estimands beyond conventional statistical or probabilistic inference.
  • Conclusion: Causal supervised learning improves predictive generalization by learning invariant features or mechanisms that reduce reliance on spurious associations.Open problems include targeted invariance benchmarks, links to adversarial robustness and meta-learning, and additional supervision signals.
  • Conclusion: Causal generative modeling enables principled controllable generation from interventional distributions and sample editing from counterfactual distributions.Open questions concern abstraction levels, scaling assignment learning to larger graphs, and when counterfactual data augmentation is effective.
  • Conclusion: Causal explanations account for causal structure when explaining predictions through feature attributions or contrastive explanations.Open problems include unifying the methodologies, scaling explanations, robustness to distribution shifts, and security and privacy against attackers.
  • Conclusion: Causal fairness develops criteria for assessing fairness and mitigating harmful disparities with respect to causal relationships in underlying data.These criteria rely on counterfactual or interventional distributions, with open questions about alternatives to equality, broader settings, hidden confounding, and social categories.
  • Conclusion: Causal reinforcement learning incorporates decision-environment structure, with claimed benefits including deconfounding, intrinsic rewards, and data efficiency; modality-specific principles target vision, NLP, and graph representation learning.The review also covers eight causal reinforcement-learning categories and benchmarks containing ground-truth causal interventions and counterfactuals.

192 Conclusion · Appendices

The conclusion highlights CausalML’s benefits while identifying untestable assumptions and gaps in benchmarks, software, and comparisons with non-causal methods. It also situates the field alongside prior CausalML surveys and machine-learning approaches for inferring causal estimands.

  • 192 Conclusion: CausalML methodology offers benefits that motivate its use across the field.
  • 192 Conclusion: Untestable assumptions remain a central issue for CausalML methodology.
  • 192 Conclusion: The field suffers from a lack of benchmarks for evaluating CausalML methods.
  • 192 Conclusion: CausalML research is constrained by a lack of software implementations.
  • 192 Conclusion: Existing work often lacks comparisons between CausalML methods and non-causal methods.
  • 192 Conclusion: Related work includes other CausalML surveys and research using machine learning to infer causal estimands.
Loading 2206.15475v3…