PaperlayerView exact revision on arXiv

Source-linked AI summary

Bias and Debias in Recommender System: A Survey and Future Directions

Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, Xiangnan He

arXiv:2010.03240v2cs.IR

TL;DR

Recommender-system research must account for biases in observational user behavior data, but existing studies are fragmented and use inconsistent terminology. This survey organizes seven bias types and existing debiasing methods, then identifies open challenges and future directions. It emphasizes that biases can interact and intensify through recommendation feedback loops, while practical debiasing remains constrained by issues such as propensity estimation and unavailable sensitive attributes.

  • Problem

    User behavior data is observational and affected by multiple biases, while research on recommendation bias is fragmented and defines bias inconsistently.

  • Method

    The survey summarizes seven bias types, provides causality-based explanations, develops a taxonomy of debiasing methods, and discusses their strengths, weaknesses, challenges, and future directions.

  • Results

    The survey shows that recommendation biases can interact, intensify data imbalance, and reinforce future user behaviors through feedback loops.

  • Takeaways & Limitations

    Understanding and addressing recommendation bias requires considering interacting biases and the feedback loop rather than treating isolated bias types alone.

  • Takeaways & Limitations

    Many fairness methods assume sensitive user, item, or group attributes are available, an assumption that may not hold in real-world scenarios.

Abstract

from arXiv · show

While recent years have witnessed a rapid growth of research papers on recommender system (RS), most of the papers focus on inventing machine learning models to better fit user behavior data. However, user behavior data is observational rather than experimental. This makes various biases widely exist in the data, including but not limited to selection bias, position bias, exposure bias, and popularity bias. Blindly fitting the data without considering the inherent biases will result in many serious issues, e.g., the discrepancy between offline evaluation and online metrics, hurting user satisfaction and trust on the recommendation service, etc. To transform the large volume of research models into practical improvements, it is highly urgent to explore the impacts of the biases and perform debiasing when necessary. When reviewing the papers that consider biases in RS, we find that, to our surprise, the studies are rather fragmented and lack a systematic organization. The terminology ``bias'' is widely used in the literature, but its definition is usually vague and even inconsistent across papers. This motivates us to provide a systematic survey of existing work on RS biases. In this paper, we first summarize seven types of biases in recommendation, along with their definitions and characteristics. We then provide a taxonomy to position and organize the existing work on recommendation debiasing. Finally, we identify some open challenges and envision some future directions, with the hope of inspiring more research work on this important yet less investigated topic. The summary of debiasing methods reviewed in this survey can be found at \url{https://github.com/jiawei-chen/RecDebiasing}.

1 INTRODUCTION

Recommender-system research faces pervasive biases because user behavior is observational, unevenly exposed, and shaped by feedback loops. This survey addresses fragmented terminology and literature by organizing bias types, debiasing methods, and future challenges.

  • Ubiquity of Biases in RS: User behavior data is observational and confounded by system exposure and user self-selection.Uneven item popularity and feedback loops further reinforce biased training data and recommendations over time.
  • Increasing Importance of Biases in RS Research: Research on recommendation biases has increased significantly since 2015, with dedicated sessions, tutorials, and multiple best-paper recognitions.Figure 1 tracks publications by year and venue, while the passage reports growing attention in major recommendation and information-retrieval venues.
  • Necessity of this Survey: Existing studies are fragmented because bias terminology is vague, inconsistent, and sometimes absent even when relevant biases are addressed.The same phenomenon may receive different names, while the same term can carry different meanings across publications.
  • Contributions: The survey summarizes seven bias types, reviews debiasing methods in a taxonomy, and discusses their strengths, weaknesses, open challenges, and future directions.It also provides causality-based explanations for data biases to clarify their nature.
  • Papers Collection: The survey collects over 180 papers through searches of major venues and citation-graph traversal for work focused on bias in recommender systems.The search covers publications from 2010 to 2021 using recommendation-related and bias-related keywords.

2 PRELIMINARIES: RECOMMENDER SYSTEM AND FEEDBACK LOOP

The recommendation lifecycle is modeled as a feedback loop among users, data, and the recommendation model. Its stages include collecting user-generated data and learning user preferences from historical interactions to predict item adoption.

  • Feedback Loop in Recommendation: The recommendation lifecycle is abstracted as a feedback loop among User, Data, and Model.The loop is presented as a bird’s-eye view of recommendation-system operation.
  • User→Data (Collection): User→Data represents collecting user-item interactions and side information such as profiles, item attributes, and contexts.This is the collection stage of the lifecycle.
  • Data→Model (Learning): Data→Model represents learning recommendation models from collected data to infer user preferences and predict target-item adoption.The learning stage derives preference from historical interactions and estimates how likely a user is to adopt an item.

2.2 Recommendation Task Formulation

This section formalizes recommendation as learning from observational user–item feedback, whose distribution can differ from the ideal test distribution. It introduces data bias and its main forms, then situates biases within the recommendation feedback loop.

  • 2.2 Recommendation Task Formulation: Recommendation learns a model f from observed user–item–label triplets to minimize prediction error in the serving stage.Feedback may be explicit, such as ratings, or implicit, such as clicks, purchases, and views.
  • 2.2 Recommendation Task Formulation: The empirical training risk estimates true risk without bias only when training and test data are identically and independently distributed.The ideal test distribution is unavailable, so learning optimizes empirical risk on the collected dataset.
  • 3.1.1 Definition of data bias.: Data bias occurs when the training-data distribution differs from the ideal test-data distribution, potentially directing model learning toward the wrong optimum.Even a sufficiently large training set can leave a gap between empirical and true-risk optima when the distributions differ.
  • 3.1.1 Definition of data bias.: The survey categorizes data bias into selection, conformity, exposure, and position bias, and analyzes each using definitions, characteristics, and causal explanations.These biases arise because user-interaction data are observational rather than experimental.
  • 3.1.2 Selection Bias.: Selection bias arises because users choose which items to rate, making observed ratings nonrepresentative and missing not at random.Survey evidence reports that users tend to rate items they like and especially very good or very bad items.
  • 3.1.2 Selection Bias.: The causal graphs distinguish desired preference effects from observation mechanisms involving exposure, public opinion, and item position.For implicit feedback, only positive interactions are observed, so unobserved interactions are used as surrogate negatives.
  • 3.1.3 Exposure Bias.: Exposure bias makes unobserved interactions ambiguous because an item may be irrelevant or simply unseen by the user.Exposure can depend on prior recommendation policies, user search, user background, and item popularity.
  • 3.1.5 Position Bias.: Inductive bias comprises model assumptions that support generalization, such as inner-product embeddings, neural networks, adaptive negative sampling, or binary-code ranking.Efficiency-oriented choices can improve learning speed or recommendation efficiency while changing the loss or reducing representation ability.

3.3 Bias and Unfairness in Results

Recommendation results exhibit popularity bias and unfairness in addition to data and model biases. Popularity amplification can reduce personalization and fairness, while imbalanced representation can systematically disadvantage groups.

  • 3.3 Bias and Unfairness in Results: Popularity bias occurs when popular items are recommended more frequently than their observed popularity would warrant.Long-tailed interaction data can make models over-score popular items and treat unpopular items as negative.
  • 3.3 Bias and Unfairness in Results: Ignoring popularity bias decreases personalization and serendipity, particularly for users who prefer niche items.The paper links repeated recommendation of popular items to poorer user experience for those users.
  • 3.3 Bias and Unfairness in Results: Popularity bias also decreases recommendation fairness by reducing visibility for high-quality items that are not popular.Popular items are not necessarily the best matches or highest-quality items.
  • 3.3 Bias and Unfairness in Results: Unfairness is systematic discrimination against individuals or groups based on intrinsic or acquired traits.Unequal representation by race, gender, age, education, or wealth can be learned from imbalanced data.
  • 3.3 Bias and Unfairness in Results: Examples include gender-related bias in recommendations and preferences for books by male authors.The survey also describes social recommendation systems reinforcing majority advantages and limiting minority reach.

3.4 Feedback Loop Amplifies Biases

Biases in recommender systems can reinforce one another through a feedback loop, while debiasing methods address these problems through distinct strategies with important limitations.

  • Feedback-loop amplification: Recommendation feedback loops amplify popularity and position biases: prominent items receive more traffic, increasing their future ranking prominence in a rich-get-richer cycle.Simulations also show popular items becoming more popular and non-popular items less popular.
  • Feedback-loop amplification: Data bias can intensify itself through data imbalance, as conformity bias grows when users increasingly behave like major groups.Selection bias and exposure bias are likewise affected by data imbalance.
  • Debiasing methods: Debiasing methods are classified by the biases they address and the method types they adopt, with reviewed methods summarized alongside their strengths and weaknesses.
  • Selection-bias debiasing: Joint generative models address selection bias by modeling both users’ rating values and their decisions about which items to rate.These models are explainable and can be effective in some scenarios, but are highly complex to train and rely on distributional assumptions that may fail.
  • Selection-bias debiasing: Model-based imputation methods learn pseudo-ratings, including asymmetric tri-training, but their performance depends on obtaining a strong pretrained predictor from biased data.The survey presents model-based imputation as a promising direction despite this dependency.
  • Selection-bias debiasing: Inverse propensity scoring uses observation probabilities as weights, yielding an unbiased estimator only when the true propensities are available or correctly specified.The survey notes that improper propensity specification leaves the estimator biased.

4.2 Methods for Conformity Bias

The supplied passages describe debiasing approaches for conformity and exposure-related effects, including social-factor modeling, propensity weighting, sampling, and exposure-based models. They also note that heuristic confidence weighting is difficult to calibrate, while propensity methods depend on accurate scores and can have high variance.

  • Conformity bias: Conformity-bias methods model public opinion or social influence to separate conformity effects from users’ underlying preferences.Approaches either incorporate behavioral features such as prior ratings and average ratings or treat ratings as synthetic outcomes of preference and social influence.
  • Exposure bias: Exposure-bias methods address the fact that users see only part of the item set, so unobserved interactions are not necessarily negative signals.The survey discusses corrections for both model training and evaluation.
  • Propensity-based methods: Inverse propensity weighting can correct selection or exposure bias by down-weighting commonly observed interactions and up-weighting rare ones.The same principle is applied to implicit-feedback evaluation and model-training objectives.
  • Weighting and sampling: Heuristic weighting and sampling assign confidence according to observability, popularity, interaction frequency, or other factors, but may not capture real negative signals.Examples include uniform negative sampling and oversampling popular negative items because they are more likely to have been exposed.
  • Limitations: Heuristic confidence weights are difficult to set because optimal confidence can vary across user-item pairs, and coarse weights can introduce empirical bias.Selecting such weights may require substantial expertise or computational resources, especially at large scale.
  • Exposure-based models: Exposure-based models introduce exposure variables so that noninteraction can reflect either nonexposure or a user’s preference after exposure.These models estimate how likely a user was exposed to an item and generate interaction behavior conditional on exposure and preference.
  • Limitations: Propensity-based estimators are theoretically unbiased for true risk, but their performance depends on propensity accuracy and inverse propensities can produce high variance.The survey states that these methods usually outperform weighting strategies, while noting that joint learning and clipping only partly mitigate the limitations.

4.5 Universal Solutions for Various Data Biases

Universal debiasing methods aim to address multiple biases and their combinations, but the reviewed approaches generally require a small unbiased dataset. Obtaining such data is difficult and costly because random logging policies can reduce user experience and platform revenue.

  • Scope of existing methods: Most existing debiasing methods target one or two biases in a specific scenario, limiting their coverage when real data contain multiple biases.The survey reports a small number of universal approaches that address multiple bias types or their combinations.
  • Universal solutions: Universal solutions transfer knowledge from a small unbiased dataset to a target recommendation model.The cited approaches include domain adaptation and related strategies for leveraging unbiased data.
  • Limitations and open problems: Obtaining unbiased data requires intervention with a random logging policy, which can hurt user experience and platform revenue.Consequently, developing universal debiasing without unbiased data remains an open problem.

4.6 Methods for Popularity Bias

The survey organizes popularity-bias methods around regularization, adversarial learning, causal graphs, re-ranking, and additional side-information strategies. These approaches seek more balanced recommendations by reducing popularity-related effects in outputs, training, exposure, or user behavior.

  • Regularization: Regularization methods guide recommendation models toward more balanced outputs by reducing relationships between item popularity and recommendation scores.Examples include Laplacian, mean-match, and Pearson-correlation regularizers.
  • Regularization: Process-oriented regularization attributes weak long-tail recommendations to insufficient training caused by sparse interaction records.The cited approach adapts training across popular and unpopular items rather than only balancing final recommendation results.
  • Adversarial learning: Adversarial learning uses a min-max game in which an adversary signals the recommender to improve recommendation opportunities for niche items.The discriminator learns associations between popular and niche items, while the recommender learns to capture niche items related to user history.
  • Causal graphs: Causal-graph methods use counterfactual intervention, back-door adjustment, or propensity scores to mitigate popularity effects in exposure, conformity, prediction, or training.The survey describes several causal sources of popularity bias and corresponding interventions.
  • Other methods: Other strategies reduce popularity bias by incorporating side information or re-ranking candidate lists with item popularity considered explicitly.The re-ranking approach first produces a relatively large list and then adjusts its ordering.
  • Fairness-aware methods: Fairness-aware ranking methods rebalance recommendation exposure through multiple ranking functions or post-processing for group-fair top-K results.The cited examples target individual equity-to-attention fairness and group fairness criteria.
  • Feedback-loop mitigation: Uniform-data methods break feedback loops by replacing normal recommendation with random item delivery, but this can reduce user experience and platform revenue.Because of this cost, uniform policies are usually restricted to a small percentage of online traffic.

5.2 General Debiasing Framework

Existing debiasing methods typically target only one or two biases, while real-world recommendation data can contain multiple biases simultaneously. The survey identifies IPS-based methods as a promising direction for general debiasing across bias types.

  • Existing debiasing methods usually address only one or two specific biases.
  • Real-world recommendation data can mix conformity, selection, popularity, and unfairness-related biases.
  • A general debiasing framework is needed to handle mixtures of biases, beginning with combinations of two or three biases.
  • IPS and its variants are presented as a promising basis for handling different kinds of biases.

5.3 Better Evaluation

Unbiased recommender-system evaluation remains difficult because current approaches depend on accurate propensity scores or substantial unbiased data. The survey calls for evaluators combining large-scale biased data with small-scale unbiased data and for standardized popularity-bias and fairness benchmarks.

  • Existing unbiased evaluation methods either require accurate propensity scores or considerable unbiased data.
  • Uniform data provides gold-standard unbiased information, but its small scale makes thorough evaluation difficult because of high variance.
  • New evaluators should combine large-scale biased data with small-size unbiased data.
  • Popularity bias and unfairness produce inconsistent score reporting, limiting comparisons among existing methods.
  • The survey proposes benchmark datasets and standardized evaluation metrics for popularity bias and unfairness.

5.4 Knowlege-enhanced Debiasing

The survey highlights auxiliary information and knowledge graphs as tools for improving debiasing. Knowledge graphs can connect user and item attributes to infer exposure information relevant to correcting exposure bias.

  • Auxiliary information can improve the efficacy of debiasing.
  • Knowledge graphs connect user and item attributes rather than treating those attributes as isolated.
  • Connections among watched movies, directors, and genres can reveal items a user was highly likely to have known.
  • This inferred exposure information is important for correcting exposure bias.
  • The survey presents knowledge graphs as a tool for developing feature-enhanced general debiasing frameworks.

5.5 Explanation and Reasoning with Causal Graph

Causal graphs offer a framework for reasoning about bias causes, effects, interventions, and counterfactual recommendation outcomes. The survey connects this causal approach with debiasing and explanation, proposing better causal models as a future direction.

  • Cause graphs can elucidate causal relationships, predict action effects, and evaluate explanations for observed events.
  • Causal graphs can represent bias causes and effects through causal paths under mild cause assumptions and additional confounding factors.
  • Treating recommendation as an intervention motivates counterfactual reasoning about what happens when items are exposed to users.
  • Causal graphs can support both debiasing and explanations by addressing why particular items are recommended.
  • The survey proposes designing causal graphs capable of reasoning, debiasing, and explanation.

5.6 Dynamic Bias

Biases in recommender systems evolve over time as fashions, user experiences, and recommendation strategies change. Understanding this evolution is identified as an important direction for analyzing dynamic effects on recommendation.

  • Recommendation biases are dynamic rather than static, evolving as fashions, user experiences, and system strategies change.The paper gives changing clothing fashions, new items and friends, and periodic strategy updates as examples.

5.7 Double-edged Sword of Bias

Biases in recommendation can have both harmful and beneficial effects, so debiasing should account for this double-edged nature while addressing fairness–accuracy trade-offs. The survey organizes these issues and identifies directions for further research.

  • Popularity bias can reflect both conformity and item quality, making it potentially beneficial when appropriately leveraged.The survey characterizes popularity bias as a double-edged sword rather than uniformly harmful.
  • Fairness-Accuracy Trade-off: Fairness interventions may sacrifice recommendation performance when different groups are treated equally with respect to sensitive attributes.The survey calls for carefully defined fairness criteria and controllable methods to manage this trade-off.
  • Conclusions: The survey reviews more than 180 papers and summarizes seven kinds of recommendation bias with their definitions and characteristics.
  • Conclusions: It also develops a taxonomy of debiasing approaches, discusses their strengths and weaknesses, and identifies open research problems.
Loading 2010.03240v2…