Deconfounded Recommendation for Alleviating Bias Amplification

Wenjie Wang; Fuli Feng; Xiangnan He; Xiang Wang; Tat-Seng Chua

Source-linked AI summary

Deconfounded Recommendation for Alleviating Bias Amplification

Wenjie Wang, Fuli Feng, Xiangnan He, Xiang Wang, Tat-Seng Chua

arXiv:2105.10648v1May 22, 2021cs.IR

TL;DR

Imbalanced historical item distributions amplify recommendation bias, while existing ranking-level controls can trade accuracy for reduced amplification and leave its root cause unresolved. The paper proposes DecRS, which uses causal modeling and approximate backdoor adjustment with user-specific inference, and reports effectiveness across two benchmarks.

Problem
Imbalanced historical item-group distributions cause recommender systems to over-recommend majority groups, while existing fairness, diversity, and calibration controls can sacrifice accuracy and do not identify the root cause.
Method
DecRS models causal relations between user representation and prediction score, applies an approximation of backdoor adjustment, and dynamically regulates it according to user status.
Results
DecRS alleviates bias amplification and improves recommendation accuracy over backbone models in extensive experiments on two benchmarks.
Takeaways & Limitations
Causal modeling provides a recommender-system approach that addresses confounding behind bias amplification rather than only controlling ranking targets.
Takeaways & Limitations
Theoretically infinite confounder sample spaces make exact backdoor adjustment intractable, motivating the proposed approximation; the causal analysis also leaves finer-grained causal relations for future work.

Abstract

from arXiv · show

Recommender systems usually amplify the biases in the data. The model learned from historical interactions with imbalanced item distribution will amplify the imbalance by over-recommending items from the major groups. Addressing this issue is essential for a healthy ecosystem of recommendation in the long run. Existing works apply bias control to the ranking targets (e.g., calibration, fairness, and diversity), but ignore the true reason for bias amplification and trade-off the recommendation accuracy. In this work, we scrutinize the cause-effect factors for bias amplification, identifying the main reason lies in the confounder effect of imbalanced item distribution on user representation and prediction score. The existence of such confounder pushes us to go beyond merely modeling the conditional probability and embrace the causal modeling for recommendation. Towards this end, we propose a Deconfounded Recommender System (DecRS), which models the causal effect of user representation on the prediction score. The key to eliminating the impact of the confounder lies in backdoor adjustment, which is however difficult to do due to the infinite sample space of the confounder. For this challenge, we contribute an approximation operator for backdoor adjustment which can be easily plugged into most recommender models. Lastly, we devise an inference strategy to dynamically regulate backdoor adjustment according to user status. We instantiate DecRS on two representative models FM and NFM, and conduct extensive experiments over two benchmarks to validate the superiority of our proposed DecRS.

1 INTRODUCTION

The paper frames bias amplification as over-recommendation of majority-group items caused by imbalanced historical interactions, and argues that existing ranking controls can sacrifice accuracy while leaving the root cause unresolved. It proposes DecRS, a causal recommender using approximate backdoor adjustment and user-specific inference, with experiments validating effectiveness.

70% of a user's watched movies being action movies can produce 90% of recommendation slots allocated to action movies.
Existing fairness, diversity, and calibration methods reduce bias amplification through ranking objectives but can sacrifice recommendation accuracy.
The paper attributes bias amplification to the historical item-group distribution confounding user representation and prediction score.
DecRS models causal relations, approximates backdoor adjustment for large confounder sample spaces, and dynamically regulates adjustment according to user status.
DecRS is instantiated on FM and NFM and evaluated extensively on two benchmarks.

2 METHODOLOGY

This section analyzes conventional recommender systems from a causal perspective to explain bias amplification, then introduces DecRS.

The section first analyzes conventional recommender systems from a causal view.
It explains the reason for bias amplification before introducing the proposed DecRS.
The section's analysis is followed by an introduction to DecRS.

2.1 A Causal View on Bias Amplification

The causal analysis represents recommender variables and identifies historical item-group distribution as a confounder between user representation and prediction score. This confounding creates spurious preference for majority groups and makes predictions vulnerable to user-interest drift.

2.1.1 Causal Graph.: The conventional recommender causal graph contains user representation U, item representation I, historical distribution D, group-level representation M, and prediction score Y.
2.1.1 Causal Graph.: D affects U, while D and U determine M and U, M, and I determine Y.
2.1.1 Causal Graph.: Because D affects both U and Y, it confounds their relationship and produces spurious correlation in prediction.
2.1.2 Conventional RS.: A user's historical distribution d_u increases scores for items in groups with higher historical click frequency, even when item matching is otherwise equivalent.
2.1.2 Conventional RS.: Majority-group items can dominate recommendation lists, including low-quality items, narrowing user interest and harming accuracy.
2.1.2 Conventional RS.: When future group interest drifts, dependence on the historical distribution can make recommendations dissatisfying.

2.2 Deconfounded Recommender System

DecRS estimates the causal effect of user representation on prediction scores using backdoor adjustment, then approximates the adjustment to make computation tractable. The approximation is exact for linear functions and has a small upper-bounded error for common nonlinear recommender functions, especially when confounder values concentrate near their expectation.

Backdoor Adjustment: DecRS targets P(Y|do(U=u), I=i) to block the confounder’s effect on user representation before estimating prediction scores.Backdoor adjustment cuts the D→U edge and uses do-calculus to derive the adjusted expression.
Backdoor Adjustment: Backdoor adjustment averages prediction scores over possible confounder values weighted by their prior P(d), rather than conditioning on the user-specific distribution d_u.This removes dependence on d_u and prevents majority-group items from receiving high scores solely because of historical click frequency.
Backdoor Adjustment: DecRS can improve robustness and accuracy under user interest drift because training considers varied historical distributions instead of relying heavily on d_u.The method is motivated by users having diverse historical distributions over item groups.
Backdoor Adjustment Approximation: The theoretically infinite sample space of D makes exact backdoor-adjustment calculation intractable, motivating an efficient approximation.In a specific dataset, the method samples users’ historical item-group distributions to form a discrete approximation set ˜D.
Backdoor Adjustment Approximation: The approximation replaces the expectation E_d[f(·)] with f evaluated at the expected confounder distribution, reducing an extensive outer sum to one function calculation.The sampled distributions estimate the prior P(d), while f can be instantiated with an existing recommender model such as FM.
Backdoor Adjustment Approximation: Approximation error ε is zero for linear f and has a small upper bound for common nonlinear functions, especially when D concentrates around its expectation.The bound follows conditions on local and asymptotic growth of the nonlinear function.

2.3 Backdoor Adjustment Operator

The backdoor-adjustment operator computes an additional group-level user representation from the expected item-group distribution and user features. This representation can be integrated into existing recommender architectures, including FM-style and later-fusion designs.

Operator Design: The operator makes backdoor adjustment usable in existing recommenders by supplying M(¯d, u) as an additional model input.Here ¯d is the prior-weighted expected distribution over item groups.
Operator Design: M(¯d, u) combines expected group probabilities, item-group embeddings, user-feature embeddings, and feature values to form a group-level user representation.The construction uses concatenated distributions and feature values with concatenated group and user embeddings.
Operator Design: A second-order FM module computes M(¯d, u) through interactions among user and item-group representations.The item-group representation v_a is randomly initialized, and ⊙ denotes element-wise product.
Integration: The resulting representation can be directly fed into models supporting multiple feature representations or combined through later fusion when they do not.The prediction module then uses f(u, i, M(¯d, u)) to calculate the score.
Integration: Training optimizes the recommender parameters with the interaction labels and a loss function such as log loss.The label indicates whether user u interacted with item i.

2.4 Inference Strategy

DecRS uses user-specific inference to adaptively regulate backdoor adjustment according to interest drift, balancing conventional and deconfounded prediction scores.

User-specific inference: The strategy preserves stronger bias adjustment for users whose interests drift while retaining more conventional recommendation behavior when bias amplification may reflect stable preferences.Stable users may benefit from recommendations concentrated on groups they prefer, whereas drifting users require stronger adjustment.
User-specific inference: DecRS estimates user interest drift with symmetric KL divergence between two timestamp-based historical item-group distributions.Higher η_u indicates that a user more easily changes their interest distribution over item groups.
User-specific inference: The inference strategy fuses conventional RS and DecRS prediction scores to regulate the impact of backdoor adjustment.Both conditional and causal prediction scores are trained before automatic fusion.
User-specific inference: The normalized drift signal and hyper-parameter α control the balance between conventional and deconfounded predictions.α∈[0,+∞) provides additional control, and α is theoretically usually close to 0 because mitigating spurious correlation improves accuracy for most users.
Design differences: DecRS differs from conventional recommendation by modeling causal effects, adding a backdoor adjustment operator, and using user-specific inference.It models P(Y|do(U=u), I=i) rather than P(Y|U=u, I=i).

3 RELATED WORK

Prior recommendation debiasing methods target fairness, diversity, calibration, or feedback biases, but often trade recommendation accuracy for reduced bias or depend on difficult propensity estimation.

Motivation: Bias amplification can narrow user interests through feedback loops, reinforce echo chambers, and expose users to undesirable majority-group items.These effects motivate methods that address imbalance without degrading recommendation quality.
Fairness in Recommendation: Fairness methods pursue similar treatment for similar individuals or equitable exposure across item groups.The related work distinguishes individual fairness from group fairness.
Fairness in Recommendation: Calibrated recommendation re-ranks items so recommended group proportions follow a user’s historical browsing distribution, but trades off ranking accuracy and fairness.A user with 70% action and 30% romance interactions would receive the same target proportions.
Diversity in Recommendation: Diversity methods increase dissimilarity among recommended items to reduce filter bubbles, but may recommend dissatisfying items by sacrificing accuracy.Similarity can be based on item categories or embeddings.
Causal Recommendation: Causal recommendation primarily debiases user-feedback problems such as position bias, clickbait, and popularity bias.Inverse Propensity Scoring reweights samples using estimated propensity scores but relies on accurate estimation and can have high variance.

4 EXPERIMENTS

Experiments evaluate DecRS on ML-1M and Amazon-Book using accuracy and bias-amplification metrics across users and baselines. Results show that DecRS reduces bias amplification without the accuracy losses observed for several ranking-control methods, and that its inference strategy improves performance.

Experimental Settings: Experiments use ML-1M and Amazon-Book, splitting timestamp-sorted interactions into 80% training, 10% validation, and 10% testing subsets.Positive instances have rating ≥4, and training uses one randomly sampled previously unseen item as a negative instance.
Experimental Settings: DecRS is instantiated on FM and NFM and compared with Unawareness, FairCo, Calibration, Diversity, and IPS.The evaluation covers recommendation accuracy and effectiveness in alleviating bias amplification.
Overall Performance: Calibration and Diversity perform worse than vanilla FM and NFM, while Unawareness and FairCo achieve comparable performance or marginal improvements.The authors attribute IPS’s inferior performance to inaccurate and high-variance propensity estimation.
Bias Amplification: DecRS achieves lower C_KL scores than calibrated recommendation across all ML-1M user groups without hurting recommendation accuracy.The comparison includes vanilla FM/NFM, calibrated recommendation, and DecRS; Amazon-Book shows similar trends but is omitted from Figure 4.
Ablation Analysis: Disabling the inference strategy lowers performance, although DecRS without it still consistently outperforms FM and NFM, especially for users with high η_u.This supports the benefit of adaptive inference alongside backdoor-adjustment approximation.
Ablation Analysis: The element-wise-product implementation DecRS-EP remains better than FM, although it is inferior to the FM-module implementation DecRS-FM.The comparison is conducted on FM over ML-1M.

5 CONCLUSION AND FUTURE WORK

DecRS uses causal modeling and approximate backdoor adjustment to remove confounder-driven spurious correlations, while an inference strategy regulates this adjustment. Extensive experiments validate its effectiveness in alleviating bias amplification and improving recommendation accuracy, while the work identifies broader causal discovery and application directions for future study.

5 CONCLUSION AND FUTURE WORK: DecRS explicitly models causal relations and uses approximate backdoor adjustment to remove confounder-driven spurious correlations.An inference strategy regulates the impact of backdoor adjustment.
5 CONCLUSION AND FUTURE WORK: Extensive experiments validate DecRS for alleviating bias amplification and improving recommendation accuracy.
Future Work: The work identifies finer-grained causal discovery in complex recommendation scenarios as an important direction for future research.The paper notes that recommendation involves many observed and hidden variables awaiting causal discovery.
Future Work: Future studies could examine DecRS for position bias, popularity bias, filter bubbles, and echo chambers.These extensions are proposed because their causes or effects are related to imbalanced training data and bias amplification.