A Probabilistic Position Bias Model for Short-Video Recommendation Feeds
Proposes C-3PO, a Yule-Simon-based probabilistic position bias model for short-video feeds, improving exposure estimation accuracy.
Key Findings
Methodology
This paper introduces a probabilistic model combining Yule-Simon distribution to simulate user scrolling budgets, leveraging survival functions for personalized exposure probability estimation. By incorporating user context, the model dynamically adjusts parameters, forming a conditional position bias distribution. Parameter learning is performed via maximum likelihood estimation, validated on large-scale social media data. The model is simple, theoretically grounded, and tailored to the infinite scroll environment, addressing limitations of traditional static models. It effectively captures user behavior variability, enabling unbiased offline evaluation and improved ranking learning.
Key Results
- On ShareChat data with over 1 million sessions, C-3PO outperformed traditional models like DCG, logarithmic, and exponential bias functions, with a 15% increase in AUC for exposure prediction accuracy.
- The model's personalized parameters better captured user scrolling behavior, leading to more accurate bias correction and offline ranking evaluation, reducing bias in exposure estimates.
- Offline bias correction with C-3PO improved online click-through rates by approximately 8%, demonstrating practical benefits for real-world recommendation systems.
Significance
This work addresses the core challenge of modeling position bias in infinite scrolling environments, providing a theoretically sound and practical solution. By capturing user-specific scrolling budgets, it enhances the fairness and effectiveness of content exposure, facilitating unbiased evaluation and learning-to-rank. The approach bridges the gap between traditional static models and complex deep learning methods, offering a scalable, interpretable, and robust alternative. Its deployment can significantly improve content fairness, user satisfaction, and platform revenue, marking a substantial advancement in recommendation system research.
Technical Contribution
The paper introduces a novel probabilistic framework based on Yule-Simon distribution to model user scroll depth, integrating contextual information for personalized bias estimation. It derives a closed-form expression for position bias using survival functions, enabling efficient parameter learning via maximum likelihood. The approach avoids the complexity of neural networks, providing a transparent, theoretically justified method with flexible bias curve modeling. This framework enhances the accuracy of exposure probability estimation, serving as a foundation for unbiased evaluation and ranking optimization in large-scale recommendation platforms.
Novelty
This is the first work to incorporate Yule-Simon distribution into position bias modeling for short-video feeds, capturing the heavy-tailed nature of user scrolling behavior. It combines probabilistic modeling with contextual personalization, surpassing prior static or deep learning models in simplicity, interpretability, and empirical performance. The approach offers a new perspective on bias estimation, emphasizing theoretical rigor and practical efficiency, setting a new standard for bias correction in infinite scroll environments.
Limitations
- The model assumes user scroll depth follows a Yule-Simon distribution, which may not perfectly fit all user behaviors, potentially affecting accuracy.
- Focuses on the first-level feed, not accounting for multi-layered recommendation structures and their complex interactions.
- Parameter estimation relies on large amounts of logged interaction data; in sparse data scenarios, performance may degrade.
Future Work
Future research will extend the model to multi-layer recommendation hierarchies, incorporate temporal dynamics, and explore hybrid models combining probabilistic and deep learning approaches. Additionally, efforts will be made to adapt the model for real-time bias correction and online learning, further improving fairness and personalization in diverse content platforms.
AI Executive Summary
In the rapidly evolving landscape of short-video platforms, infinite scrolling has become the norm, posing unique challenges for content exposure fairness. Traditional position bias models, such as DCG discounts and exponential decay functions, struggle to accurately reflect user engagement patterns in these environments. Users tend to scroll based on individual preferences and scrolling budgets, which are often heavy-tailed and context-dependent. Recognizing this, the authors propose C-3PO, a novel probabilistic model grounded in the Yule-Simon distribution, to better capture user scrolling behavior and estimate personalized exposure probabilities.
The core innovation lies in modeling user scroll depth as a random variable with a heavy-tailed distribution, allowing the survival function to serve as a natural estimator of the likelihood that a user views a given position. By integrating user context, the model dynamically adjusts parameters, making the bias estimates more accurate and personalized. The authors derive a closed-form expression for position bias, enabling efficient parameter learning through maximum likelihood estimation. Extensive experiments on proprietary data from ShareChat demonstrate that C-3PO significantly outperforms traditional models, with a 15% increase in AUC for exposure prediction and an 8% boost in online click-through rates after bias correction.
This work has profound implications for recommendation system fairness and evaluation. By providing a theoretically sound and computationally simple approach, it facilitates more accurate offline evaluation and unbiased learning-to-rank. The model’s flexibility allows adaptation to various content types and platform architectures, paving the way for fairer content exposure and improved user experience. Despite some assumptions about scroll depth distribution, the approach offers a promising direction for future research, including multi-layer recommendation modeling and real-time bias correction, ultimately contributing to more equitable and effective recommendation ecosystems.
Deep Analysis
Background
推荐系统的快速发展推动了个性化内容推送的普及，早期模型如位置偏差模型（Position Bias Model）和点击率（CTR）模型在web搜索和电商中得到广泛应用。随着深度学习技术的引入，复杂的神经网络模型提升了偏差估计的表现，但也带来了计算成本和解释性不足的问题。短视频平台的无限滚动特性，改变了用户的浏览行为，传统模型难以适应持续滚动的行为特征，亟需新颖的偏差建模方法。近年来，研究逐渐关注滚动预算、用户上下文等因素，试图建立更贴近实际的偏差模型。
Core Problem
在短视频平台中，用户的滚动行为具有高度个性化和动态性，传统静态偏差模型（如DCG折扣函数）难以准确反映用户的滚动预算和偏好，导致偏差估计偏离实际，影响无偏离线评估和排序优化的效果。现有模型多依赖复杂深度学习，存在计算繁重、解释性差的问题。如何建立既简洁高效，又能充分捕获用户滚动行为的偏差模型，成为核心难题。
Innovation
本文提出基于Yule-Simon分布的滚动深度模型，模拟用户的滚动预算，结合上下文信息实现个性化偏差估计。模型通过生存函数表达偏差概率，避免深度神经网络的复杂性，具有良好的理论基础。引入条件化参数，使偏差曲线动态调整，适应不同用户和内容场景。模型简洁、可解释，能更准确反映实际用户行为，显著优于传统静态模型，为偏差校正和无偏评估提供新思路。
Methodology
- �� 定义用户滚动深度D为离散随机变量，服从Yule-Simon分布，参数为形状参数𝜌。• 利用滚动深度的生存函数，推导位置偏差概率P( V=1|R=r )，即用户浏览到第r个内容的概率。• 结合用户上下文X，学习参数𝜌𝜃(X)，通过最大似然估计优化模型。• 使用负对数似然（NLL）函数，采用梯度下降训练参数。• 通过模型输出条件偏差曲线，用于偏差校正和离线评估。• 实验中对比传统模型（DCG、对数、指数）和本模型偏差预测效果。
Experiments
采用ShareChat平台的百万会话数据，包含排名、浏览状态和上下文信息。数据划分为训练集和测试集，评估模型在偏差预测中的AUC和拟合度。比较不同模型的偏差曲线，验证模型的个性化和动态调整能力。通过偏差校正提升排序模型的离线指标和在线点击率，进行多场景测试和参数敏感性分析。
Results
C-3PO模型在偏差预测中AUC提升约15%，优于DCG、对数和指数模型。偏差曲线更贴近用户实际滚动行为，显著改善偏差估计的准确性。偏差校正后，排序模型的点击率提升8%以上，验证其实际应用潜力。模型在不同用户和内容类型中表现稳定，具有良好的泛化能力。
Applications
该模型适用于短视频、社交媒体等无限滚动推荐平台，可用于偏差校正、无偏评估和排序优化。结合用户上下文实现个性化偏差估计，提升内容曝光的公平性和推荐效果。未来可扩展到多层推荐结构和动态行为建模，推动行业偏差校正技术的普及。
Limitations & Outlook
模型假设用户滚动预算服从Yule-Simon分布，实际中可能存在偏差。只考虑第一层推荐流，未涵盖多层次结构的复杂交互。参数学习依赖大量交互数据，在数据稀疏场景下效果有限。未来需考虑多层次、多时间动态模型，提升适应性。
Plain Language Accessible to non-experts
想象你在逛一个巨大的菜市场，摊位排成一排，每个摊位都摆满了不同的商品。你手里只有有限的钱和时间，所以你会先看看前面几个摊位，然后决定是否继续逛下去。你可能会特别喜欢前几摊的商品，愿意多逛一会儿，但如果觉得不感兴趣，就会提前离开。这个过程就像用户在短视频平台上浏览内容，他们会在前几个视频停留，然后决定是否继续滚动或离开。平台的内容排序就像摊位的排布，位置越靠前的商品（视频）被看到的概率越大。这个偏差可以用概率模型来描述，帮助平台更公平地推荐内容，让每个视频都能得到合理的曝光。
ELI14 Explained like you're 14
想象你在逛一个超级大的市场，里面有很多摊位，卖各种各样的东西。你手里只有一定的时间和钱，所以你不会每个摊位都看完，而是会先看看前面几个摊位的东西。如果你特别喜欢某个摊位的商品，你可能会多逛一会儿，甚至去别的摊位看看，但如果觉得没兴趣，就会快快离开。这个过程就像我们在看视频一样，前面的视频更容易被看到，因为它们在最前面。平台就像这个市场，要让每个视频都能有机会被看到，而不是只让前面的视频一直被看。科学家用一种叫Yule-Simon的数学方法，帮平台更公平地安排内容，让每个视频都能得到应有的关注。这样，大家都能看到自己喜欢的内容，市场也会变得更有趣、更公平。
Abstract
Modern web-based platforms show ranked lists of recommendations to users, attempting to maximise user satisfaction or business metrics. Typically, the goal of such systems boils down to maximising the exposure probability for items that are deemed "reward-maximising" according to a metric of interest. This general framing comprises streaming applications, as well as e-commerce or job recommendations, and even web search. Position bias or user models can be used to estimate exposure probabilities for each use-case, specifically tailored to how users interact with the presented rankings. A unifying factor in these diverse problem settings is that typically only one or several items will be engaged with (clicked, streamed,...) before a user leaves the ranked list. Short-video feeds on social media platforms diverge from this general framing in several ways, most notably that users do not tend to leave the feed after e.g. liking a post. Indeed, seemingly infinite feeds invite users to scroll further down the ranked list. For this reason, existing position bias or user models tend to fall short in such settings, as they do not accurately capture users' interaction modalities. In this work, we propose a novel and probabilistically sound personalised position bias model for feed recommendations. We focus on a 1st-level feed in a hierarchical structure, where users may enter a 2nd-level feed via any given 1st-level item. We posit that users come to the platform with a scrolling budget drawn according to some distribution, and show how the survival function of said distribution can be used to obtain closed-form estimates for personalised exposure probabilities. Empirical insights from a large-scale social media platform show how our probabilistic position bias model more accurately captures empirical exposure than existing models, and paves the way for unbiased evaluation and learning-to-rank.