Unbiased Learning to Rank with Query-Level Click Propensity Estimation: Beyond Pointwise Observation and Relevance
Proposes DualIPW, combining query-level and position-level click propensity estimation, to mitigate relevance saturation bias in unbiased learning to rank.
Key Findings
Methodology
This paper introduces a dual inverse propensity weighting (DualIPW) mechanism that integrates query-level and position-level click propensity models. The query-level model h(cs) uses an LSTM to encode click sequences and predict overall query click propensity, addressing relevance saturation bias. The position-level model g(k) adapts from DLA to estimate position bias. The combined loss function optimizes a ranking model f by weighted cross-entropy, ensuring unbiased learning. Theoretical proofs guarantee unbiasedness, validated by experiments on Baidu ULTR dataset showing superior performance over SOTA baselines.
Key Results
- On Baidu ULTR data, DualIPW improves nDCG@10 by over 0.5% and ERR@10 by 0.6% compared to baselines like DLA and UPE. It performs especially well on low-frequency queries, demonstrating robustness in sparse data scenarios.
- Ablation studies confirm that combining query-level and position-level components yields significant gains, validating the dual bias correction approach. Click weight analysis indicates more accurate reflection of user relevance, especially at top ranks.
- Results show that the model effectively reduces relevance saturation bias, leading to fairer and more accurate rankings. It outperforms existing methods in real-world complex bias environments, confirming the theoretical advantages.
Significance
This work advances bias correction in learning to rank by explicitly modeling query-level relevance saturation bias, a previously underexplored aspect. It bridges the gap between theoretical unbiasedness and practical robustness, enabling search engines to produce fairer, more accurate results. The approach addresses real-world complexities, making it highly relevant for large-scale industrial applications, and sets a new benchmark for bias mitigation in IR systems.
Technical Contribution
The paper's key technical contribution is the formulation of a query-level click propensity model h(cs) based on listwise click sequence divergence, integrated with position bias g(k). The dual weighting scheme ensures unbiased learning under complex bias conditions. Theoretical proofs establish unbiasedness, and the model architecture facilitates efficient training. This systematic integration of listwise and pointwise bias correction is novel and enhances robustness over previous single-level approaches.
Novelty
This is the first work to incorporate query-level click propensity into ULTR, addressing relevance saturation bias explicitly. Unlike prior methods focusing solely on position bias, the dual mechanism captures the global query bias, providing a more comprehensive bias correction framework. The listwise modeling of click sequences introduces a new perspective, significantly improving real-world performance.
Limitations
- The model assumes that click probability depends on relevance, observation, and query-level propensity, which may oversimplify individual user behaviors. It may not fully capture personalized biases or temporal dynamics. Computational complexity increases with sequence modeling, requiring substantial training resources. Handling extreme bias scenarios remains challenging, necessitating further refinement.
Future Work
Future directions include integrating user personalization features, modeling temporal bias variations, and extending to multi-task learning frameworks to balance relevance and fairness. Exploring more efficient sequence encoding methods and real-time bias adaptation will further enhance practical deployment. Additionally, applying this approach to other modalities like recommendation systems could broaden its impact.
AI Executive Summary
In the realm of information retrieval, unbiased learning to rank (ULTR) aims to produce fair and accurate search results despite inherent biases in user click data. Traditional models primarily focus on position bias, assuming that clicks are solely influenced by the position of results. However, real-world user behavior is more complex, often involving relevance saturation bias, where users examine multiple relevant results but click only a few due to limited patience or satisfaction. This discrepancy leads to false negatives and biased training signals, hampering model performance.
To address this, the authors propose a novel DualIPW framework that combines query-level and position-level click propensity estimation. The core innovation is a listwise query propensity model h(cs), which encodes click sequences using an LSTM to estimate the overall likelihood of a query receiving clicks, capturing relevance saturation effects. Simultaneously, a position bias g(k) model adjusts for position-dependent biases, inspired by DLA. The combined loss function applies inverse weighting at both levels, ensuring the learned ranking model is unbiased.
Theoretical analysis confirms that DualIPW can learn an unbiased ranking function under complex bias conditions. Extensive experiments on Baidu's large-scale ULTR dataset demonstrate significant improvements over existing methods like DLA, UPE, and IOBM, especially on low-frequency queries. Results show over 0.5% gain in nDCG@10 and 0.6% in ERR@10, validating the effectiveness of incorporating query-level bias modeling.
This approach has broad implications for real-world search engines, enabling fairer, more robust ranking systems. It addresses the critical challenge of relevance saturation bias, which has been underexplored, and provides a comprehensive framework adaptable to various bias scenarios. Future work will explore personalization, temporal bias modeling, and multi-modal applications, further pushing the frontier of unbiased learning to rank.
Deep Analysis
Background
近年来，学习排序（LTR）技术在搜索引擎中取得巨大成功，但偏差问题始终困扰其性能。早期研究多基于位置偏差模型（PBM），假设用户点击由位置和相关性共同决定，采用逆倾向加权（IPW）校正偏差。代表性工作如DLA、UPE、IOBM等，试图通过模型估计偏差概率，缓解偏差影响。然而，真实场景中用户行为更复杂，偏差来源多样，包括用户耐心、信息满足度等，导致模型在实际应用中表现不佳。尤其在大规模真实数据中，偏差的多层次、多维度特性未被充分捕捉，限制了偏差校正的效果。
Core Problem
核心问题在于现有ULTR方法多依赖位置偏差假设，忽视用户行为的多样性和查询整体偏好，导致偏差估计不足。尤其是在复杂偏差环境下，模型易受假阴性和偏差累积影响，难以学习真实的排序关系。如何结合查询层面整体偏好信息，准确估计点击倾向，成为提升无偏排序的关键。该问题难点在于偏差的多源、多尺度特性，以及大规模数据的高效建模需求。
Innovation
本研究的创新点包括：1）引入查询级点击倾向模型，利用点击序列的全局信息，估算用户对整个结果列表的偏好；2）提出双重逆倾向加权机制，将查询级与位置级偏差同时校正，增强模型鲁棒性；3）采用序列模型（LSTM）编码点击序列的对数比，提升偏差估计的准确性。这些创新突破了传统只关注位置偏差的局限，系统性缓解了相关偏差，显著改善了模型在真实场景中的表现。
Methodology
- �� 输入：查询-文档对特征、点击序列。
- �� 查询级偏差模型：利用LSTM编码点击序列的对数比，预测查询整体点击倾向h(cs)。
- �� 位置级偏差模型：借鉴DLA思想，估计每个位置的偏差g(k)。
- �� 逆倾向加权：结合模型输出，计算无偏损失函数，优化排序模型f。
- �� 训练过程：交替优化查询偏差模型和排序模型，确保偏差估计的准确性。
- �� 目标：最大化无偏排序性能指标（如nDCG）同时减缓偏差影响。
Experiments
使用百度ULTR大规模搜索会话数据，包含14个特征，筛选出10+结果的会话。对比基线包括BM25、Naive、IPW、DLA、UPE、IOBM等。指标采用nDCG@10和ERR@10，进行5次随机种子平均。调参方面，采用AdamW优化器，学习率在2e-6到6e-6之间，训练2轮，批次大小30。模型在低频查询上表现尤为优越，验证了偏差缓解的有效性。
Results
DualIPW在nDCG@10上提升0.5%以上，ERR@10提升0.6%以上，明显优于DLA、UPE等。消融实验显示，查询级与位置级结合优于单一机制。低频查询中性能提升最明显，点击权重分析表明模型更合理调整偏差，反映用户真实行为。模型在真实偏差环境中表现优越，验证了理论假设。
Applications
该方法适用于搜索引擎、推荐系统等场景，能有效缓解偏差，提高排序公平性和用户满意度。实现条件包括丰富的点击数据和特征信息，模型训练成本较高，但可显著改善偏差影响。未来可结合个性化特征，提升模型适应性。
Limitations & Outlook
模型假设偏差由相关性、观察概率和查询偏好共同决定，未充分考虑用户个性化差异。对极端偏差和稀疏点击场景效果有限，训练成本较高，参数调优复杂。未来需结合用户画像，提升偏差估计的个性化和效率。
Plain Language Accessible to non-experts
想象你在一家厨房里做饭，厨师想让每道菜都变得完美。可是，厨房里有个问题：某些食材总是被优先放在前面，厨师容易忽略后面的食材。为了让每个食材都能被公平对待，厨师设计了一个特殊的规则，考虑每个食材的“偏爱程度”和“摆放位置”。这样，他可以确保每个食材都能得到合理的烹饪时间，不会因为位置偏见而被忽略。这个规则就像论文中的DualIPW机制，既考虑整体偏好（查询偏好），也考虑单个位置的偏差，让菜肴更公平、更美味。它帮助厨师（搜索引擎）更好地理解每个食材（文档）的价值，从而做出更合理的选择。
ELI14 Explained like you're 14
想象你在学校图书馆找书，很多书都摆在不同的架子上。有些书放在显眼的位置，大家都能看到；有些书藏在角落，可能没人注意到。你会不会只看前面几排就觉得所有书都差不多了？其实，很多好书可能在后面，只是因为位置不好看不到。这个论文就像在帮图书馆设计一种聪明的办法，让每本书都能被公平看到，不会因为摆放位置偏见而被忽略。它用一种特别的“魔法”——结合你对整个书架的偏好和每个位置的偏差，让你找到真正喜欢的书。这样一来，无论书放在哪里，都能被公平地推荐给你，让你找到最喜欢的那本。
Abstract
Most existing unbiased learning-to-rank (ULTR) approaches are based on the user examination hypothesis, which assumes that users will click a result only if it is both relevant and observed (typically modeled by position). However, in real-world scenarios, users often click only one or two results after examining multiple relevant options, due to limited patience or because their information needs have already been satisfied. Motivated by this, we propose a query-level click propensity model to capture the probability that users will click on different result lists, allowing for non-zero probabilities that users may not click on an observed relevant result. We hypothesize that this propensity increases when more potentially relevant results are present, and refer to this user behavior as relevance saturation bias. Our method introduces a Dual Inverse Propensity Weighting (DualIPW) mechanism -- combining query-level and position-level IPW -- to address both relevance saturation and position bias. Through theoretical derivation, we prove that DualIPW can learn an unbiased ranking model. Experiments on the real-world Baidu-ULTR dataset demonstrate that our approach significantly outperforms state-of-the-art ULTR baselines. The code and dataset information can be found at https://github.com/Trustworthy-Information-Access/DualIPW.