PaperlayerView exact revision on arXiv

Source-linked AI summary

Causal Inference in Recommender Systems: A Survey and Future Directions

Chen Gao, Yu Zheng, Wenjie Wang, Fuli Feng, Xiangnan He, Yong Li

arXiv:2208.12397v2cs.IR

TL;DR

Correlation-based recommender systems do not fully represent the causal processes behind user decisions and system interactions. This survey synthesizes causal-inference research for recommendation, organizing methods around data bias, missing data, noise, interpretability, and related challenges. It also identifies unresolved limitations, including reliance on predefined causal graphs and the limited explainability of graph neural network recommenders.

  • Problem

    Correlation-based recommender systems do not capture causal user decisions and strategy-driven interactions, while recommendation data can contain bias and noise.

  • Method

    The survey systematically reviews causal-inference approaches for recommendation and organizes them by the challenges causal methods address.

  • Results

    The survey synthesizes how causal inference addresses recommendation data bias and supports causal modeling across existing research.

  • Takeaways & Limitations

    Causal inference is presented as a framework for mitigating shortcomings in non-causal recommendation models and motivating further research.

  • Takeaways & Limitations

    Existing approaches often rely on potentially inaccurate, simplistic predefined causal graphs, while graph neural network recommenders remain difficult to explain.

Abstract

from arXiv · show

Recommender systems have become crucial in information filtering nowadays. Existing recommender systems extract user preferences based on the correlation in data, such as behavioral correlation in collaborative filtering, feature-feature, or feature-behavior correlation in click-through rate prediction. However, unfortunately, the real world is driven by causality, not just correlation, and correlation does not imply causation. For instance, recommender systems might recommend a battery charger to a user after buying a phone, where the latter can serve as the cause of the former; such a causal relation cannot be reversed. Recently, to address this, researchers in recommender systems have begun utilizing causal inference to extract causality, thereby enhancing the recommender system. In this survey, we offer a comprehensive review of the literature on causal inference-based recommendation. Initially, we introduce the fundamental concepts of both recommender system and causal inference as the foundation for subsequent content. We then highlight the typical issues faced by non-causality recommender system. Following that, we thoroughly review the existing work on causal inference-based recommender systems, based on a taxonomy of three-aspect challenges that causal inference can address. Finally, we discuss the open problems in this critical research area and suggest important potential future works.

1 INTRODUCTION

Recommender systems traditionally model correlations in behavioral, feature, and feature-behavior data, although recommendation interactions are shaped by causal processes. This survey organizes causal-inference research around the data problems it addresses and reviews open directions in the area.

  • Recommender systems commonly learn preferences by fitting historical behaviors alongside user, item, and contextual information.
  • Correlation-based recommendation overlooks causal user decisions and strategy-driven interactions, including the possibility that non-exposure explains an unobserved user-item interaction.
  • Existing correlation-driven systems face data bias, including observational-data distortions such as conformity and popularity bias.
  • Causal recommendation methods address bias through causal graphs, missing data through counterfactual reasoning, and interpretability through more controllable models.
  • Prior surveys separately cover recommender systems or causal inference, leaving no comprehensive survey focused on causality-driven recommender systems.
  • The survey systematically reviews causal recommendation research by explaining why causal inference is needed, how it enhances recommendation, and which unresolved problems remain.

2 BACKGROUND

The survey introduces background knowledge and fundamental concepts from causal inference and recommender systems before reviewing their integration.

  • The survey first introduces background knowledge and fundamental concepts from causal inference and recommender systems.

2.1 Causal Inference

Causal inference formalizes cause–effect relationships and estimates causal effects using structural causal models or potential outcomes. Its practical use includes interventions, counterfactual reasoning, adjustment methods, randomized experiments, observational estimation, and causal discovery.

  • Structural Causal Models: Structural causal models represent variables and their causal relationships with directed acyclic graphs and structural functions.Graphs encode relationships through nodes and directed edges; each structural function computes a variable from its parent nodes.
  • Structural Causal Models: Interventions set a variable to a chosen value while blocking the effects of its parents, enabling causal reasoning about controlled changes.The do-calculus operation do(X = x) removes incoming paths to X and forces X to equal x.
  • Structural Causal Models: Counterfactual inference estimates what an outcome would have been if a treatment had taken a different value than its observed value.Counterfactuals address treatment settings that do not occur in the factual world.
  • Potential Outcome Framework: The potential outcome framework estimates treatment effects without requiring a causal graph, but each individual reveals only one potential outcome.Individual treatment effects therefore cannot be directly calculated and are extended to population-level average treatment effects.
  • Causal Effect Estimation: Causal effects can be estimated through randomized experiments or observational-data methods such as backdoor adjustment, frontdoor adjustment, and inverse propensity weighting.Backdoor adjustment addresses observed confounders, frontdoor adjustment can address unobserved confounders under its criterion, and inverse propensity weighting uses treatment probabilities.
  • Causal Discovery: Causal discovery constructs causal graphs from data when prior causal knowledge is unavailable, using independence tests, score-based algorithms, and machine learning approaches.The survey identifies causal effect estimation and causal discovery as two important research topics.

2.2 Recommender System

Recommender systems learn from user–item interactions and related features to predict future behavior. Their main model families are collaborative filtering and click-through-rate prediction, using embeddings, feature interactions, and point-wise or pair-wise objectives.

  • Overview: Recommendation models use historical user–item interactions and contextual information to learn preferences and predict future interactions.The recommendation output is a model that maps users and items to predicted interactions.
  • Collaborative Filtering: Collaborative filtering includes matrix-factorization, neural-network, and graph-neural-network models that represent users and items with latent embeddings.Matrix factorization measures interaction likelihood through embedding similarity, while later models extend the similarity function or propagate graph information.
  • Collaborative Filtering: Graph neural recommendation models perform message passing over the user–item bipartite graph to inject high-order connectivity into similarity measurements.Neighborhood aggregation uses interacted items for users and interacting users for items.
  • Click-Through Rate Prediction: Click-through-rate prediction extracts high-order interactions among feature fields after transforming raw features into embeddings.Feature embedding matrices represent optional features, with user and item identity commonly occupying the first two fields.
  • Click-Through Rate Prediction: Feature-interaction modules can use inner products, multilayer perceptrons, or stacked self-attention layers.These designs underlie models such as FM, DeepFM, and AutoInt.
  • Optimization: Recommendation objectives are commonly point-wise or pair-wise, with Logloss used for interaction prediction and BPR encouraging separation between positive and negative samples.The pair-wise objective compares positive interactions with negative samples.

Causal Inference

The survey organizes causal inference concepts and causal-inference-based recommendation around model architecture, data augmentation, causal-effect estimation, and debiasing. It frames these approaches as causality-enhanced recommendation methods addressing issues in non-causal models.

  • Taxonomy: The survey’s causal-inference recommendation taxonomy includes model architecture, data augmentation, causal-effect estimation, and debiasing.These categories are presented as causality-enhanced directions for recommender systems.
  • Motivation: Figure 4 illustrates three typical issues of non-causality recommendation models and how causal inference addresses them.The section introduces the benefits of causal inference from three aspects.

3 WHY CAUSAL INFERENCE IS NEEDED FOR RECOMMENDER SYSTEMS

Causal inference is needed because recommendation data can be biased, incomplete, noisy, and poorly representative of user preferences. These limitations motivate causal methods that model data-generation processes and use counterfactual reasoning.

  • Data bias: Recommendation data can be biased because historical interactions and item attributes do not faithfully reflect user preferences.Interaction bias includes popularity, category imbalance, conformity, and exposure effects; attribute bias can encourage shortcut learning.
  • Data bias: Causal modeling identifies confounders in recommendation-data generation and mitigates their spurious effects on estimated preferences.For example, item popularity can affect both exposure and interaction probability, producing popularity bias.
  • Data missing: Limited feedback and missing user or item features prevent observed data from fully representing user interests.Users interact with only a small fraction of candidate items, and feature collection can be costly.
  • Data missing: Traditional approaches to missing data use prediction-error correction or observation-probability weighting, but they have limited accuracy and generalization.Causal inference provides descriptions of how recommendation data are generated, supporting alternative treatments of missingness.
  • Data noise: Recommendation data contain natural and malicious noise, including inaccurate feedback, collection errors, fake ratings, and fabricated orders.Such noise can arise from users’ imprecise preference judgments, preprocessing choices, or adversarial manipulation.
  • Data noise: 40% RMSE differences and 9.56%-21.81% Recall@20 and 3.92%-8.81% NDCG@20 drops demonstrate that noisy data can substantially degrade recommendation performance.These comparisons are reported against noiseless or cleaned data in prior experiments.

3.3 Beyond-accuracy Concerns in Recommender Systems

Beyond accuracy, recommender systems face explainability, diversity, filter-bubble, and fairness concerns. Causal inference is presented as a way to reason about biased feedback, user behavior, and counterfactual fairness.

  • Explainability: Explainability requires understanding why items are recommended or omitted, supporting transparency and trustworthiness.The paper distinguishes explainable recommendation models from explainable recommendation results.
  • Diversity and filter bubble: Similarity-based recommendation can produce homogeneous content and contribute to filter bubbles that narrow users’ views and radicalize ideas.Biased feedback loops can further reinforce homogeneous exposure and distort subsequently collected data.
  • Diversity and filter bubble: Causal inference can support exploration beyond observed categories by modeling why users consume items and reasoning about missing or biased feedback.This may help recommendation systems address the accuracy-diversity dilemma within the scope described by the survey.
  • Fairness: Recommendation fairness involves both personalized user-side concerns and equitable treatment of items or item attributes.Users may face different demographic fairness concerns, while items may be overlooked during recommendation.
  • Fairness: Causal fairness analysis can formulate questions about how recommendations would differ under counterfactual changes to group membership.The survey frames fairness as a counterfactual-reasoning problem.
  • Scope: The survey systematically reviews how causality-enhanced recommendation methods partially address these limitations.It presents recent advances after discussing why causal inference is needed.

4 TECHNICAL DETAILS OF EXISTING WORKS OF CAUSAL INFERENCE-BASED RECOMMENDER SYSTEMS

The survey organizes causal-inference work in recommendation around three major problems created by correlation-only modeling. Figure 5 provides the overall illustration of these existing approaches.

  • Overview: Existing causal-inference methods for recommendation are organized around three major issues of correlation-based recommendation models.The survey refers readers to Figure 5 for the overall illustration and then introduces the details.

4.1 Causal Inference-based Recommendation for Addressing Data Bias

Causal debiasing methods address confounding, collider, and counterfactual effects through structural causal models and potential-outcome techniques. The surveyed approaches include adjustment, propensity-based estimation, doubly robust learning, and path-specific counterfactual inference.

  • Taxonomy: Causal debiasing methods are grouped into confounding-effect, collider-effect, and counterfactual-inference approaches.This taxonomy organizes representative methods for addressing data bias in recommendation.
  • Confounding effect: Structural causal models address confounding through backdoor or frontdoor adjustment, depending on whether relevant confounders can be observed.Backdoor adjustment blocks effects of observed confounders, while frontdoor adjustment is used when observed-confounder assumptions are infeasible.
  • Confounding effect: Backdoor adjustment has been applied to popularity bias and category-imbalance amplification by treating exposure-related variables as confounders.These methods adjust recommendation training or user representations to reduce over-recommendation of popular or majority categories.
  • Potential-outcome methods: Inverse propensity weighting estimates causal effects from exposure probabilities, but propensity estimation can be difficult and high-variance.Doubly robust estimators augment propensity weighting with an error-imputation model, while experimental data can improve estimation but are costly and sparse.
  • Collider effect: Collider structures arise when multiple causes affect observed interactions, so conditioning on interactions can create correlations between user interests and conformity.Causal inspection is used to motivate methods for mitigating conformity bias.
  • Counterfactual inference: Counterfactual inference estimates path-specific effects by comparing factual behavior with a counterfactual world that removes biased user or item features.This approach has been used to remove exposure-feature effects such as attractive titles in clickbait mitigation.
  • Scope of surveyed methods: Table 2 summarizes representative causal-inference methods addressing data missing and data noise.The table’s scope includes recommender systems and causal inference methods.

4.2 Causal Inference-based Recommendation for Addressing Data Missing and Noise

Causal inference-based recommendation addresses sparse, noisy, and policy-dependent interaction data by generating counterfactual evidence, correcting rewards, and estimating recommendation effects. The surveyed methods cover counterfactual augmentation, denoising, uplift estimation, and off-policy evaluation across several recommendation settings.

  • Problem setting: The section frames missingness as both limited engagement relative to the candidate pool and unavailable outcomes for items not recommended.These conditions make user-preference learning difficult when models rely on incomplete and noisy training data.
  • Data augmentation: Counterfactual augmentation generates interactions under conditions different from factual data to address sparse observations.It considers alternative recommendation lists, interaction sequences, feature settings, and distributions.
  • Counterfactual generation: A simulator trained on factual interactions produces counterfactual outcomes after inputs are intervened or randomized.The pipeline consists of modeling the data-generation process, intervening on inputs, and inferring outputs with the simulator.
  • Data denoising: Causal methods address noisy feedback by reweighting early rewards and leveraging sparse reliable feedback.Importance sampling modifies rewards when delayed outcomes are unavailable, while colliding inference uses reliable feedback such as ratings.
  • Causal effect estimation: Recommendation effects are estimated through uplift modeling and off-policy evaluation because rewards for unselected actions are often unavailable.The surveyed work applies potential-outcome methods, sampling, and reinforcement-learning perspectives to estimate these effects.

4.3 Beyond-accuracy RecSys with Causal Inference

Causal inference extends recommender systems beyond accuracy by supporting explainability, diversity, filter-bubble reduction, and fairness. The surveyed approaches use counterfactual reasoning, causal graphs, causal discovery, user control, and adversarial learning to model these objectives.

  • Explainable recommendation: Causal inference improves explainability by modeling how causes lead to recommendation outcomes rather than relying only on correlations.The surveyed explainability methods include counterfactual learning, causal graph-guided representation learning, and causal discovery.
  • Explainable recommendation: Counterfactual methods explain recommendations by contrasting factual and alternative scenarios or identifying changes that reverse an outcome.CountER also uses causal discovery to extract relations from historical interactions and recommended items.
  • Diversity: Causal inference can support more diverse recommendations and reduce filter bubbles by explicitly modeling causal effects of user-decision factors.One framework adds user-controlled commands and counterfactual user representations to seek out-of-bubble content.
  • Fairness: Counterfactual fairness evaluates whether recommendations change when a user-profile feature is altered, with filtered embeddings used to predict sensitive features adversarially.The approach combines adversarial prediction with recommendation losses.

5 OPEN PROBLEMS AND FUTURE DIRECTIONS

The survey identifies predefined causal structures, changing distributions, opaque graph models, and data-driven simulators as important open challenges. It calls for causal discovery, robustness, improved evaluation, tighter online-offline integration, causal GNNs, and causality-supported simulators.

  • Causal discovery: Predefined causal graphs may be inaccurate, omit hidden causes, and oversimplify the many factors involved in real-world user decisions.The survey therefore emphasizes causal-discovery methods that learn relations from recommendation data.
  • Robustness: Recommendation models must remain robust across rapidly changing data distributions, domains, interfaces, and objectives.Examples include newly registered users, new products, multiple app tabs, and diverse recommendation settings.
  • Causality-aware GNNs: Causality-aware GNN research still needs to explain model behavior and explicitly connect message passing with causal reasoning.Current GNN-based recommenders can remain black boxes, while their graph modules and causal inference components are often insufficiently integrated.
  • Simulators and evaluation: Data-driven recommendation simulators may lack causal foundations, motivating causal reinforcement-learning methods for estimating rewards with unobserved confounders.The survey also highlights long-term utility evaluation, offline-online bridging, and causality-supported environments as future directions.

6 CONCLUSION

The survey presents causal inference as an important emerging direction for recommender systems and systematically reviews why it applies to shortcomings of non-causal methods. It aims to motivate both current researchers and newcomers to the field.

  • Survey contribution: The paper provides a comprehensive and systematic survey of causal inference-based recommender-system literature.It examines the rationale for applying causal inference and how it addresses shortcomings of non-causal recommendation models.
  • Research impact: The survey is intended to motivate active researchers and inspire researchers considering work in this emerging area.Its stated audience includes both established participants and prospective entrants.
Loading 2208.12397v2…