Unbiased Learning-to-Rank with Biased Feedback

Learning to rank (LTR) has evolved from supervised methods relying on explicit relevance judgments (e.g., RankSVM, LambdaRank) to leveraging implicit feedback like clicks. Early works focused on modeling relevance directly, but high annotation costs limited scalability. Click models such as Position-Based Model (PBM) and Cascade Model attempted to correct biases but depended on query repetition and complex parameter estimation. Recent advances incorporate causal inference, using propensity scoring to mitigate bias effects. Despite progress, existing methods often require query duplication or system randomization, which are impractical in many real-world settings, especially with long-tail queries and personalized data. The need for a theoretically sound, scalable, and query-agnostic bias correction framework remains urgent.

The core challenge is that implicit feedback data is inherently biased—clicks are influenced by position, presentation, and user exposure, not solely relevance. This bias skews the training data, leading to models that favor results at the top of rankings and ignore less visible but relevant items. Traditional methods either rely on query repetition, active system perturbation, or heuristic heuristics, which are costly or unreliable. The key problem is developing a bias correction approach that works with observational data, without query duplication, and provides theoretical guarantees of unbiasedness, robustness to noise, and scalability in diverse scenarios.

The paper's main innovations include: 1) Applying causal inference via counterfactual estimation to derive unbiased ranking performance estimators; 2) Using click models as propensity estimators rather than relevance sources; 3) Designing a propensity-weighted SVM-Rank that integrates IPS into the optimization process; 4) Developing randomized intervention strategies to calibrate propensity parameters with minimal system disruption; 5) Theoretically proving consistency and robustness under click noise and model misspecification. These innovations collectively enable bias correction from observational data, bypass query repetition, and extend applicability across various bias types and real-world conditions.

• �� Establish a counterfactual inference framework to model true relevance and biases; • Use click models (e.g., PBM) to estimate click propensities, considering position and trust biases; • Collect observational data, including presented rankings, clicks, and partial relevance signals; • Implement randomized interventions (e.g., swapping results at different ranks) to calibrate propensity parameters; • Calculate inverse propensity scores (IPS) for each clicked result, adjusting for presentation bias; • Incorporate these IPS weights into a modified SVM-Rank objective, penalizing misrankings proportionally to propensity-adjusted relevance; • Optimize the resulting convex quadratic program efficiently for large datasets; • Validate the approach through synthetic experiments and live search engine deployment.

Synthetic datasets simulated with controlled bias levels and click noise tested the method’s robustness, showing >20% improvement over biased models. Real-world logs from a commercial search engine involved A/B testing, comparing standard and bias-corrected models, with CTR gains of 15% and ranking quality improvements of 10%. Hyperparameters like regularization C and propensity calibration intervals were tuned via cross-validation. Ablation studies examined the impact of propensity estimation accuracy, noise levels, and intervention strategies, confirming the method’s stability and scalability across diverse scenarios.

Results demonstrated that bias correction via IPS significantly outperforms uncorrected models, with a 25% increase in NDCG and 15% CTR improvements. The approach maintained robustness under click noise and model misspecification, outperforming existing methods like result randomization and heuristic bias correction. The method scaled linearly with data size, supporting deployment in large-scale industrial systems. Ablation experiments highlighted the importance of accurate propensity estimation and intervention calibration, guiding future refinements.

This bias correction framework applies directly to search engines, recommendation systems, and personalized advertising, particularly where query repetition is sparse. It enables models to learn from observational logs without system perturbation, improving relevance and fairness. The approach supports online learning, adaptive bias calibration, and can be integrated into existing ranking pipelines with minimal disruption.

Dependence on click model accuracy; misestimated propensities can impair bias correction. The initial intervention experiments add complexity. In scenarios with extreme bias or sparse data, performance may degrade. Extending to multi-objective or multi-modal feedback remains challenging, requiring further research.