A Comprehensive Survey on Trustworthy Recommender Systems

Authors: Wenqi Fan, Xiangyu Zhao, Xiao Chen, Jingran Su, Jingtong Gao, Lin Wang, Qidong Liu, Yiqi Wang, Han Xu, Lei Chen, Qing Li

Venue: ACM Computing Surveys (to appear), 2022 — arXiv:2209.10117

TL;DR

Recommender systems, while beneficial to users, can amplify harmful content, introduce discrimination, and enable adversarial attacks. This survey provides a comprehensive framework for trustworthy recommender systems across six critical dimensions: Safety & Robustness (defense against adversarial attacks), Non-discrimination & Fairness (equitable treatment of users and items), Explainability (transparent decision-making), Privacy (protection from leakage), Environmental Well-being (energy efficiency), and Accountability & Auditability (responsibility and transparency). The work systematizes taxonomies, methods, applications, and future research directions for each dimension.

Contributions

The survey introduces and formalizes six interdependent dimensions of trustworthiness in recommender systems:

1. Safety & Robustness — methods to detect and defend against adversarial attacks that manipulate recommendation outcomes (e.g., fake users, poisoned ratings, adversarial training techniques).

2. Non-discrimination & Fairness — taxonomies of bias sources (data bias, model bias, feedback loop bias) and fairness definitions (procedural, outcome, causal, personalized, explainable fairness) with corresponding evaluation metrics (Absolute Difference, Variance, Min-Max Difference, Entropy, KL-Divergence).

3. Explainability — methods for making recommendation mechanisms interpretable through model-intrinsic and post-hoc explanations.

4. Privacy — techniques to prevent private data leakage from recommendation models.

5. Environmental Well-being — measurement of energy consumption and techniques for sustainable, efficient recommendation.

6. Accountability & Auditability — mechanisms for responsibility, answerability, and sanctionability in recommender system governance.

Method

The survey is organized around these six dimensions, each presented as a comprehensive taxonomy:

• Adversarial attack and defense perspective (Safety & Robustness): attacks categorized by adversary goal (target attacks to promote/demote items vs. general attacks to degrade quality), attack vector (data poisoning, model manipulation, inference-time perturbations), and defense strategies (detection-based and adversarial robust training).

• Bias and fairness taxonomy (Non-discrimination & Fairness): sources of bias mapped to three stages of the recommendation pipeline (data collection, model training, online serving); fairness definitions span from causal fairness to personalized fairness; evaluation metrics quantify disparities in recommendation quality and item exposure.

• Explainability mechanisms: model-intrinsic approaches (attention, feature importance) and post-hoc explanations to increase user trust and system transparency.

• Privacy threats and defenses: privacy attacks (reconstruction, membership inference) and privacy-preserving techniques (differential privacy, federated learning).

• Sustainability metrics: measurement of GPU power consumption, proposal of energy-aware recommendation techniques.

• Governance frameworks: auditability, responsibility assignment, and sanctionability in multi-stakeholder systems.

For each dimension, the survey reviews representative state-of-the-art algorithms, cites practical applications (healthcare drug recommendations, e-commerce ranking, online jobs), and identifies open problems.

Results

The paper does not present new empirical results but rather a comprehensive systematic review of the field. It identifies key challenges:

• Most research focuses on one or two dimensions while ignoring interactions among dimensions (e.g., privacy constraints may conflict with fairness goals).
• Real-world deployment of trustworthy recommender systems remains limited despite methodological advances.
• Evaluation of trustworthiness remains fragmented across disconnected metrics and datasets.

Open research directions include:

• Developing multi-objective optimization to balance competing trustworthiness goals.
• Designing fairness constraints that account for dynamic user preferences and item popularity shifts.
• Creating standardized benchmarks and toolkits for evaluating trustworthy recommendation.
• Integrating trustworthiness considerations into collaborative, multi-stakeholder designs.
• Studying cross-domain trustworthiness principles.

Connections

• Related to Propagation-based fake news detection and information spread — recommender systems amplify both quality content and false information, making robustness critical in news recommendation contexts.
• Overlaps with Content moderation — both aim to control harmful content distribution.
• Connected to Fairness and Bias In ML — fairness in algorithmic systems is foundational to trustworthy recommendation.
• Cited alongside A Comprehensive Survey on Trustworthy Graph Neural Networks: Privacy, Robustness, Fairness, and Explainability as complementary perspectives on trustworthiness in graph-based systems.

Notes

Strengths: The breadth of coverage across six interconnected dimensions makes this a rare single-source reference for the full scope of trustworthiness in recommender systems. The taxonomy of adversarial attacks and defenses is rigorous. The emphasis on multi-stakeholder perspectives and practical applications grounds the work in real deployments.

Weaknesses: The survey does not systematically address interactions or trade-offs among dimensions — e.g., adding privacy-preserving noise may degrade fairness or accuracy. Limited discussion of how trustworthiness requirements differ across domains (news, healthcare, finance vs. e-commerce). No quantitative empirical comparison of defense methods across a unified benchmark.

Fit for this wiki: Relevant to misinformation because recommender systems are a primary vector for false-information amplification in social media and news platforms. Understanding robustness, fairness, and explainability in recommender systems is essential for designing systems that resist manipulation and reduce the spread of misinformation. This survey bridges the gap between general recommender-system research and information-quality concerns.