Causal Inference in Recommender Systems: A Survey and Future Directions

Introduction

• Correlation-based recommender systems operate in feedback loops, yet causal user decisions and policy-driven exposure mean observed interactions can be biased, missing or noisy, and inadequate for beyond-accuracy goals.The introduction identifies fairness, explainability, transparency, and diversity as objectives that can conflict with accuracy.Source: §1 · pp. 2–3
• Causal inference determines and leverages cause-effect relations from experimental or observational data; the survey introduces potential-outcome and structural causal models as foundational frameworks [128],,.The potential-outcome framework calculates treatment effects, while structural causal models represent causal relationships with graphs and equations.Source: §1 · p. 2
• Causally aligned, interpretable, and controllable models can support explainability, diversity, and fairness; causal discovery and disentangled representations are among the techniques described.The survey says interpretable models can be manipulated to improve diversity and steered to secure fairness across user demographics.Source: §1 · p. 4
• For bias, causal graphs treat popularity or exposure as confounders for backdoor adjustment and conformity as a collider; counterfactual reasoning augments missing data, while IPW estimates recommendation-model effects.The survey groups these approaches as data debiasing, augmentation, and denoising.Source: §1 · pp. 3–4
• The survey fills a gap left by prior recommender-system and causal-inference surveys, and differs from the bias-focused survey through its broader scope and emphasis on causal techniques and other challenges.The authors note that the overlap with is small; that survey treats causal inference as one of several bias methods and examines bias types and origins in greater depth.Source: §1 · p. 4
• The authors present a pioneering systematic survey of causal inference for recommendation, covering work through 2023 and proposing unresolved problems and promising future directions.They organize prior work around why causal inference is needed and how it can enhance recommendation.Source: §1 · pp. 3–5

2 BACKGROUND

• The survey first presents background knowledge and fundamental concepts of causal inference and recommender systems.Source: §2 · p. 5

2.1.1 Structural Causal Models.

• SCMs represent causal relationships with DAGs and structural functions, including chain, fork, and collider structures that distinguish mediation, confounding, and conditioning.In the recommender examples, item quality can confound price and preference, while ignoring it can produce spurious associations.Source: §2.1.1 · p. 5
• Interventions use do(X = x) to block X’s parents and set its value, while counterfactuals estimate outcomes under treatment values not observed in reality.For example, an intervention can set an item’s price, and a seller can consider potential sales if they had purchased advertising.Source: §2.1.1 · pp. 5–6

2.1.2 Potential Outcome Framework.

• The potential outcome framework estimates treatment effects without a causal graph; because each individual reveals only one potential outcome, ITE is unobservable and ATE summarizes effects across a population.For binary treatment, ITE is Y_i^1 − Y_i^0; Pearl describes the framework as logically equivalent to SCMs, though it does not reason over causal graphs.Source: §2.1.2 · p. 6

2.1.3 Causal Effect Estimation and Causal Discovery.

• Randomized experiments enable direct treatment-control comparisons, but cost or user-experience and profitability risks make observational-data estimation important in recommender systems.With a causal model such as an SCM, the causal estimand is translated into a statistical estimand and estimated from observed data while distinguishing causal effects from confounding associations.Source: §2.1.3 · p. 7
• Backdoor adjustment blocks treatment-outcome backdoor paths using eligible covariates to address observed confounders, while frontdoor adjustment uses a mediator to address possible unobserved confounders.For high-dimensional adjustment sets, propensity scores and inverse propensity weighting offer an alternative to directly modeling the full covariate set.Source: §2.1.3 · p. 7
• Causal discovery constructs a causal graph from data when prior relationships are unavailable, using conditional-independence tests, score-based methods, or machine-learning approaches such as Zhu et al.’s reinforcement-learning method [142].Traditional tests may rely on faithfulness; score-based algorithms [34, 87] compare graph quality with observed data, and a survey reviews discovery methods.Source: §2.1.3 · pp. 7–8

2.2.1 Overview.

• Historical user–item interactions are used to model preferences and predict the likelihood of future interactions.The model takes interaction records as input and produces a function that scores a user–item pair.Source: §2.2.1 · p. 8

2.2.2 Recommendation Model Design.

• CTR prediction uses user and item features, embedding raw fields and learning their high-order interactions.Feature-interaction modules include inner products in FM, multilayer perceptrons in DeepFM [24], and stacked self-attention in AutoInt.Source: §2.2.1 · p. 8, §2.2.2 · p. 9
• CF scores user–item pairs through latent embeddings, with models spanning matrix factorization, neural networks, and graph neural networks.MF uses an inner product; NCF [31] uses an MLP to model nonlinearity; NGCF passes messages over the user–item graph to incorporate high-order connectivity.Source: §2.2.2 · pp. 8–9

2.2.3 Objective Function.

• Recommendation optimization uses pointwise Logloss to predict interactions or pairwise BPR loss to increase the score gap between positive and negative samples.The pairwise objective compares positive interactions (y_u,i = 1) with negative samples (y_u,j = 0).Source: §2.2.3 · pp. 9–10

3 WHY CAUSAL INFERENCE IS NEEDED FOR RECOMMENDER SYSTEMS.

• The section frames causal inference’s benefits for recommender systems through three aspects, illustrated in Fig. 4.Source: §3 · p. 10

3.1.1 Data bias in recommender systems.

• Historical interactions can overrepresent popular items, producing popularity bias, or frequent categories, causing models to neglect other preferences.These biases can be amplified through feedback loops, contributing to unfairness and filter bubbles.Source: §3.1.1 · p. 10
• User behavior can reflect conformity as well as preference, while exposure algorithms also shape which feedback is collected.The passage identifies both as sources of biased interaction data.Source: §3.1.1 · pp. 10–11
• Clickable item attributes can become shortcuts: attractive video titles or covers may prompt clicks without reflecting content preference.Such shortcuts can undermine recommendation quality and leave systems vulnerable to attacks that exploit those features.Source: §3.1.1 · p. 11

3.1.2 The necessity of causal inference for data debiasing.

• Causal inference traces recommendation bias to data-generation processes, including backdoor confounding that can inflate scores for items beyond what they deserve.For example, popularity can affect both prior exposure and interactions through conformity, causing collaborative filtering to over-recommend popular items.Source: §3.1.2 · p. 11
• Bias can also arise when observed interactions diverge from true preference matching because item attributes directly affect interaction status.This is distinct from confounding through a shared cause of inputs and interactions.Source: §3.1.2 · p. 11
• Debiasing requires modeling the causal effect of inputs X on interactions Y rather than their correlation, treating X as treatment and Y as outcome.The causal effect concerns how Y changes when X is forcibly changed from a reference status to its observed value.Source: §3.1.2 · p. 11

3.2.1 Data missing in recommender systems.

• Interactions are constrained by deployed recommendation policies, users may give feedback selectively, and costly feature collection can leave user and item data missing.Because observed data cannot fully represent users’ interests, existing recommendation methods can produce sub-optimal results.Source: §3.2.1 · pp. 11–12

3.2.2 The necessity of causal inference for data missing.

• Earlier methods computed errors for missing ratings or weighted observed ratings by collection probability, while causal descriptions of data generation can improve accuracy and generalization.The survey argues that causal inference can provide prior knowledge to data-driven models and alleviate the negative impact of missing data.Source: §3.2.2 · p. 12

3.2.3 Data noise in recommender systems.

• Noise can be natural or malicious, arising from inaccurate preference judgments, rating discretization, delayed or misleading implicit feedback, mistaken clicks, or adversarial ratings and fake orders/comments.Natural-noise sources include inaccurate user impressions or collection errors, preference/rating mismatch, delayed labels, and gaps between implicit feedback and satisfaction.Source: §3.2.3 · p. 12

3.2.4 The necessity of causal inference for data denoising.

• Noisy-trained recommendations dropped 9.56%-21.81% on Recall@20 and 3.92%-8.81% on NDCG@20 versus cleaned-data training, while noisy versus noiseless data differed by about 40% in RMSE.Only 60% of users kept the same rating when re-rating a movie, and experiments found statistically significant MAE differences between collaborative-filtering models trained on original and new ratings.Source: §3.2.4 · pp. 12–13

3.3.1 Explainability.

• Causal inference is central to explainability because correlation-based explanations can mislead, while explanations should connect model components and prediction scores or decision factors and user behavior.Explainable recommendation includes model explanations and result explanations; aspect-aware methods [11], including Wang et al., are among existing approaches, but their limitations are noted.Source: §3.3.1 · p. 13

3.3.2 Diversity and Filter Bubble.

• Similarity-based recommendations can reinforce homogeneous exposure and biased feedback, while causal inference can mitigate data bias and support recommendations beyond existing categories to address the accuracy-diversity dilemma.The section identifies the filter bubble as narrowing views and argues that understanding causal relationships in user behavior can help meet user demands with items outside familiar categories.Source: §3.3.2 · pp. 13–14

3.3.3 Fairness.

• Fairness spans user- and item-side concerns, and counterfactual causal reasoning evaluates whether recommendations would change if a user’s group or an item’s feature were different.Association-based methods [28] can be inadequate [45], while causal studies examine how outputs evolve with inputs; item-side approaches include unbiased learning and heuristic ranking adjustments.Source: §3.3.3 · p. 14

• Existing causal-inference recommendation work is organized around three major issues in correlation-only models, as illustrated in Fig. 5.Source: §4 · p. 14, Figure 5 · p. 15

4.1.1 Confounding Effect.

• SCM-based backdoor adjustment blocks confounder paths; Zhang et al. use it for popularity bias, while Wang et al. approximate it for category-driven bias amplification.Backdoor adjustment intervenes on treatment variables to cut confounder paths to outcomes.Source: §4.1.1 · pp. 15–16
• When confounders are unobserved, frontdoor adjustment is explored for global and personalized confounders [125], while other work addresses selection bias and unmeasured confounders with Robust Deconfounder.Zhu et al. [144] analyze conditions for applying frontdoor adjustment in recommendation.Source: §4.1.1 · p. 16
• Potential-outcome methods use inverse-propensity weighting (IPW) for unbiased learning; propensity estimation evolved from separate models to multitask learning and sequential recommender USR.The propensity represents the probability of observing user-item feedback.Source: §4.1.1 · pp. 16–17
• Because propensity scores can have high variance, doubly robust estimators add imputation [27]; RCT aggregation addresses costly sparse experiments [9], while IPS can use social-network data.Experimental data such as random exposure can enhance doubly robust estimation, but RCT data are costly.Source: §4.1.1 · p. 17

4.1.2 Colliding Effect.

• Conditioning on observed interactions can make user interest and conformity appear correlated; disentangling their representations with cause-specific data mitigates conformity bias.The resulting user representations are described as more robust and interpretable.Source: §4.1.2 · p. 17

4.1.3 Counterfactual Inference.

• Counterfactual inference estimates path-specific effects by comparing factual and imagined worlds, removing biased feature effects such as exposure-driven clickbait and item-driven popularity bias.An adversarial method models counterfactual exposure and optimizes recommendations against a worst-case scenario [123].Source: §4.1.3 · pp. 17–18

4.2.1 Causal Inference for Data Missing.

• Counterfactual data augmentation follows a three-step modeling, intervention, and inference paradigm; in sequential and feature-based recommendation it considers altered interaction sequences and feature-level preferences and generates counterfactual outputs.Interventions can use random replacements or learning-based selection of hard samples and decision-boundary items or preferences.Source: §4.2.1 · pp. 19–20
• For missing or shifted-distribution data, Wang et al. use a variational auto-encoder, Mu et al. change knowledge-graph relations, and CausPref [33] learns invariant preferences.Mu et al. jointly train the counterfactual generator and recommender; CausPref combines a recommendation-specific DAG learner with anti-preference negative sampling for varied out-of-distribution settings.Source: §4.2.1 · p. 20

4.2.2 Causal Inference for Data Noise.

• Delayed feedback makes early rewards noisy; Zhang et al. use importance sampling to re-weight rewards, while Wang et al. use reliable feedback to mitigate noise under sparsity.For example, a purchase may occur days after an item is added to a cart, so its reward is unknown when early training uses the sample; colliding inference leverages sparse reliable feedback on predictions.Source: §4.2.2 · pp. 20–21

4.2.3 Causal Effect Estimation for Recommendation.

• Uplift estimates the causal effect of recommendation by distinguishing purchases induced by recommendations from purchases users would make anyway.Sato et al. use potential outcomes and ULO to learn sample-level uplift; IPW supports unbiased offline learning and online evaluation, while Xie et al. use tensor factorization and RDD.Source: §4.2.3 · p. 21
• Off-policy evaluation estimates deployed recommendation-policy rewards from historical data using model-based, propensity-based, or doubly robust estimators.For slate recommendation, Swaminathan et al. estimate policy performance from logged data; Li et al. estimate list clicks, and Mcinerney et al. use a lower-variance sequential counterfactual estimator. Kiyohara et al. [46] propose a Cascade Doubly Robust estimator under sequential top-down interaction.Source: §4.2.3 · pp. 21–22

4.3.1 Causal Inference for Explainable Recommendation.

• Causal explainability methods use counterfactuals, disentangled long- and short-term preference embeddings, and search behavior as an instrumental variable to separate causal from non-causal model parameters.CountER contrasts factual and counterfactual scenarios to identify an item whose substitution reverses the recommendation; the preference embeddings use defined causal relations and self-supervised signals.Source: §4.3.1 · p. 22
• Knowledge-graph paths selected through reinforcement learning support accurate explanations, while job-skill causal discovery and counterfactual feature changes identify skills whose learning could enable employment.The knowledge-graph method uses interaction history to extract causal relations, and the job-skill method finds the feature whose modification can lead to employment.Source: §4.3.1 · p. 23

4.3.2 Causal Inference for Improving Diversity and Alleviating Filter Bubble.

• Causal modeling of user-decision factors can improve diversity and reduce filter bubbles; methods include user-controlled counterfactual embeddings, backdoor adjustment, and dynamic causal collaborative filtering.Counterfactual embeddings remove out-of-date features while retaining accuracy; backdoor adjustment blocks imbalanced item-category effects, and dynamic causal collaborative filtering estimates post-intervention preferences to mitigate echo chambers while maintaining strong recommendation performance.Source: §4.3.2 · p. 23

4.3.3 Causal Inference for Fairness in Recommendation.

• Counterfactual fairness seeks unchanged recommendation-probability distributions when a feature changes, using filtered user embeddings and adversarial prediction of sensitive features.A filtering module removes sensitive-feature information after the embedding layer; a prediction module then uses the filtered embeddings to predict sensitive features alongside the primary recommendation losses.Source: §4.3.3 · pp. 23–24

5.1 Causal Discovery for Recommendation

• Causal discovery is needed because predefined graphs may be inaccurate or too simplistic to capture the many factors shaping real-world decisions.Constraint-based methods such as PC [93] and FCI [92], and GES methods, face computational and robustness challenges at scale; deep-learning [42] and reinforcement-learning [142] approaches are emerging.Source: §5.1 · p. 24

5.2 Causality-aware Stable and Robust Recommendation

• Causality-aware models may help recommendations remain stable across changing data distributions, domains, scenarios, and offline-to-online deployment.Prior research [57] highlights their potential for domain adaptation and out-of-distribution generalization.Source: §5.2 · p. 24

5.3 Causality-aware Graph Neural Network-based Recommendation

• Future work should explain GNN-based recommendations and explicitly connect GNN message passing with causal inference and reasoning.GNN-based models can extract high-order graph similarity, but often remain black boxes; causality-aware models that include GNN modules have not fully joined those modules with causal inference.Source: §5.3 · pp. 24–25

5.4 Causality-aware Simulator and Environment for Recommendation

• Causality-aware reinforcement learning could support simulators and environments that model dynamic, sequential user-system interactions.Existing data-driven simulators use reinforcement and imitation learning [39] [36] but may lack causal grounding; causal interventions can estimate rewards under unobserved confounding, while causal bandits offer theoretical performance bounds over non-causal bandits [49, 126]. Future work also includes evaluating long-term utility and bridging offline and online settings.Source: §5.4 · p. 25

6 CONCLUSION

• The survey systematically examines why causal inference applies to recommendation and how it mitigates shortcomings of non-causal models.Source: §6 · p. 25
• The authors present the survey as an initial stride toward comprehensively reviewing causal inference research in recommender systems.Source: §6 · p. 25
• The paper aims to motivate researchers already active in the field and inspire those considering starting research in this area.Source: §6 · p. 26