Ensure Timeliness and Accuracy: A Novel Sliding Window Data Stream Paradigm for Live Streaming Recommendation
Proposes Sliver sliding window data stream paradigm, boosting live recommendation timeliness and accuracy, with a 6.76% CTR increase demonstrated.
Key Findings
Methodology
This paper introduces the Sliver paradigm, which employs a shortened, continuously sliding window to generate training samples, balancing label timeliness and accuracy. The approach integrates a time-sensitive re-reco strategy that periodically re-requests recommendation services, reducing latency between request and impression. Experiments on a real-world multi-task dataset from Kuaishou platform validate the method, employing models such as Shared Bottom, MMOE, ESSM, CGC, and PLE. Results show consistent improvements across all metrics, with an average CTR increase of 6.76% and NFN increase of 3.69%, outperforming traditional fixed-window schemes.
Key Results
- Offline evaluations reveal that Sliver outperforms fixed-window baselines across four models, with target-specific improvements ranging from 3% to 8%. The average CTR gains are 6.76%, and NFN gains are 3.69%. Label accuracy at 30 seconds post-impression exceeds 85%, significantly higher than longer window schemes.
- Online A/B testing confirms the offline findings, with CTR improving by 6.76% and new followers increasing by 3.69%. The system maintains stable performance under dynamic content changes, demonstrating practical deployment feasibility.
- The combination of sliding window sampling and re-reco request strategy effectively shortens delays, enhances feature timeliness, and adapts to content dynamics, leading to better user engagement.
Significance
This work pioneers from a data stream perspective to address the core challenge of balancing timeliness and label accuracy in live streaming recommendation systems. The sliding window mechanism and re-reco strategy provide a new paradigm for real-time data processing, significantly improving the responsiveness of recommendation models in rapidly changing environments. The approach bridges the gap between theoretical data stream design and practical deployment, offering a scalable solution for industry applications. It advances the state-of-the-art in online learning and real-time recommendation, with broad implications for content platforms, e-commerce, and online education, ultimately enhancing user experience and platform engagement.
Technical Contribution
The main technical innovation lies in the design of the sliding window data stream, which continuously slides a short window (e.g., 30 seconds) to generate training samples, effectively reducing label delay while maintaining accuracy. The integration of a time-sensitive re-reco mechanism allows dynamic adjustment of recommendation requests, further minimizing latency. The approach is compatible with multiple multi-task learning architectures, demonstrating high flexibility. Theoretically, it offers a new perspective on data stream management, balancing differential (timeliness) and integral (accuracy) principles, inspired by calculus. Empirically, extensive experiments validate its superiority over fixed-window methods in both offline and online settings.
Novelty
This is the first comprehensive proposal of a sliding window data stream paradigm tailored for live recommendation systems, explicitly addressing the trade-off between label timeliness and accuracy. Unlike prior works focusing solely on model architectures or content features, this approach innovatively redefines data stream design itself. The continuous sliding mechanism, combined with a dynamic re-request strategy, distinguishes it from traditional fixed-window or delayed feedback solutions, offering a fundamentally new way to handle content volatility and user behavior delays in real-time environments.
Limitations
- The approach requires high computational resources due to frequent sliding window updates and re-reco requests, which may challenge scalability in massive systems.
- Optimal window size and re-reco frequency depend on content dynamics and platform latency, necessitating adaptive tuning for different scenarios.
- Current experiments are primarily conducted on the Kuaishou platform; cross-platform generalization and long-term stability need further validation.
Future Work
Future research will explore adaptive window sizing based on content change rates, integrating reinforcement learning to optimize re-reco schedules dynamically. Additionally, multi-modal content understanding can be incorporated to enrich feature representations, further improving recommendation quality. Extending the framework to other domains like e-commerce and online education, and optimizing computational efficiency, are promising directions to make the approach more scalable and universally applicable.
AI Executive Summary
In the era of mobile internet, live streaming has become a dominant form of real-time content interaction, attracting millions of users worldwide. However, the dynamic and rapidly changing nature of live content presents significant challenges for recommendation systems, which must deliver timely and accurate suggestions to maintain user engagement. Traditional recommendation approaches rely on fixed window data streams, which involve a trade-off: longer windows ensure comprehensive labels but introduce latency, reducing the system’s responsiveness; shorter windows improve timeliness but risk generating noisy or incomplete labels.
This paper addresses the fundamental issue by proposing a novel data stream paradigm called Sliver, inspired by principles of calculus. Instead of fixed, static windows, Sliver employs a continuously sliding short window (e.g., 30 seconds), which moves incrementally over time, generating training samples in near real-time. This approach inherently guarantees the timeliness of labels, as samples are produced promptly after user behaviors occur. To further enhance the system’s responsiveness, the authors introduce a time-sensitive re-reco strategy, which periodically re-requests recommendations, reducing delays caused by network latency and request-response cycles.
The authors validate their approach through extensive offline experiments on a multi-task dataset collected from Kuaishou, one of China's largest live streaming platforms. They compare Sliver against traditional fixed-window schemes of varying sizes (one hour, five minutes), employing multiple multi-task learning models such as Shared Bottom, MMOE, ESSM, CGC, and PLE. Results demonstrate that Sliver consistently outperforms these baselines across all key metrics, with an average CTR improvement of 6.76% and NFN increase of 3.69%. The label accuracy at 30 seconds post-impression exceeds 85%, significantly higher than longer window schemes.
Furthermore, deploying Sliver in a real-world setting on Kuaishou’s platform, the authors conducted online A/B tests. The results confirmed offline findings, with CTR gains of 6.76% and increased user engagement metrics, validating the practical effectiveness of their method. The combined use of sliding window sampling and dynamic re-requesting effectively reduces delays, improves feature timeliness, and adapts to content volatility.
This work marks a significant advancement in live recommendation system design, shifting the focus from model-centric solutions to data stream architecture. By balancing the differential (timeliness) and integral (accuracy) aspects of data processing, Sliver provides a scalable, flexible framework capable of handling the fast-paced nature of live content. Its potential applications extend beyond live streaming to any domain requiring real-time data processing, such as e-commerce, online education, and social media.
Looking ahead, future research will explore adaptive window mechanisms, multi-modal content integration, and reinforcement learning-based optimization to further enhance system robustness and efficiency. Despite some computational challenges, the proposed paradigm opens new avenues for building truly real-time, high-precision recommendation systems, promising a more engaging and responsive user experience in the digital age.
Deep Analysis
Background
随着移动互联网的快速发展，直播平台成为内容消费和互动的主要渠道之一。早期推荐系统多依赖静态内容和用户偏好，难以应对直播内容的实时变化。近年来，学术界提出多种模型（如LiveRec、ContentCTR、DRIVER）引入时间信息、多模态特征和动态用户行为，提升内容的时效性理解。工业界如快手、抖音不断优化推荐算法，结合深度学习和多任务学习，增强个性化和实时性。然而，内容的快速变化和用户行为的延迟反馈，仍然是系统性能的瓶颈。传统方案多采用固定窗口策略，存在标签滞后和时效性不足的问题，亟需从数据流设计层面提出创新方案。
Core Problem
核心问题在于如何设计一种数据流方案，在保证标签时效性的同时，确保标签的准确性。固定窗口方案在缩短窗口以提升时效性时，会引入延迟反馈和标签偏差；而长窗口虽保证了标签的完整性，但导致模型训练滞后，影响实时响应。请求与曝光之间的延迟（𝜏）和行为反馈的延迟（𝛿）共同限制了系统的实时性。如何在这两者之间找到平衡点，成为亟待解决的难题。
Innovation
本文提出的创新点在于设计滑动窗口数据流（Sliver），通过连续滑动短窗口（如30秒）生成训练样本，显著缩短标签生成延迟，确保标签的时效性。同时，结合时间敏感的re-reco策略，动态调整推荐请求频率，减少请求与曝光的时间差，提升系统响应速度。该方案融合多任务学习模型，支持多目标预测，增强模型的适应性和鲁棒性。不同于传统固定窗口的静态设计，滑动窗口实现了连续、动态的样本生成机制，有效缓解内容快速变化带来的挑战。
Methodology
- �� 定义多任务直播推荐问题，考虑用户行为、内容变化和延迟因素，建立时间轴模型。
- �� 设计滑动窗口机制，设定窗口长度（如30秒），在每个窗口结束时生成训练样本，缩短样本生成延迟。
- �� 实现连续滑动，确保样本的实时性，避免标签滞后问题。
- �� 引入时间敏感的re-reco策略，周期性请求推荐服务，减少请求与曝光的延迟。
- �� 采用多任务学习架构（如Shared Bottom、MMOE、ESSM、CGC、PLE），在多目标（点击、关注、点赞）上训练模型。
- �� 在快手平台真实多任务数据集上进行离线和在线实验，评估模型性能和系统效果。
- �� 通过AB测试持续优化参数（窗口长度、请求频率），确保系统稳定性。
Experiments
使用快手平台收集的多任务直播数据集，涵盖用户行为时间戳、内容信息和多目标预测标签。对比基线方案（如一小时窗口、五分钟窗口），采用CTR、NFN等指标进行评估。离线实验中，调节窗口参数，观察标签准确率和模型性能变化。在线AB测试中，随机分组用户，监测CTR和NFN的变化，确保统计显著性。还进行了窗口长度敏感性分析，验证滑动窗口的优越性。
Results
Sliver在所有目标任务上均优于传统固定窗口方案，平均CTR提升6.76%，NFN提升3.69%。离线指标显示，30秒窗口下标签准确率超过85%，优于长窗口方案。在线测试中，用户点击率显著提升，用户留存和互动指标也同步改善。消融实验验证了滑动窗口和re-reco策略的协同作用，确保系统在内容快速变化环境中的稳定性和高效性。
Applications
该方法适用于实时直播推荐、短视频平台、在线教育和电商直播等场景。只需平台具备实时数据采集和快速模型更新能力，即可部署滑动窗口机制，提升内容匹配速度和用户体验。未来结合多模态内容理解，有望实现更精准的个性化推荐，推动行业向更高的实时性和智能化发展。
Limitations & Outlook
当前方案对计算资源要求较高，频繁滑动和请求可能带来系统压力。窗口参数需根据内容变化速率动态调节，否则可能影响标签代表性。实验主要在快手平台验证，跨平台适应性和长时间稳定性尚待验证。未来需优化算法效率，降低成本，并结合多模态信息增强模型鲁棒性。
Plain Language Accessible to non-experts
想象你在一家工厂里，每天都在生产不同的产品。为了确保每个产品都符合标准，工厂会在一定时间内检查生产线上的产品质量。以前，工厂用一个固定的时间段，比如一小时，来检查所有产品，但这样一来，检测结果就会滞后，不能及时反映最新的生产情况。后来，工厂决定用一个更聪明的方法：每隔30秒就检查一次生产线，连续不断地监控。这样，工厂可以更快地发现问题，及时调整生产流程。这个方法就像论文中的滑动窗口机制，缩短了检测的时间，让工厂的质量控制变得更快、更准确。
ELI14 Explained like you're 14
嘿，你知道吗？在直播平台上，主播和观众之间的互动就像是在玩一个快节奏的游戏。平台要不停地推荐新内容，让你觉得每次都很新鲜，但内容变化得太快，系统就像在追赶时间一样。以前，推荐系统用一个固定的“时间箱”来决定哪些内容可以用来学习，但这个箱子太大或太小都不好。太大了，内容太旧，不能反映最新的热点；太小了，又没有足够的内容，模型学不到东西。于是，科学家们想出了一个新办法：用一个“滑动的箱子”，每隔几秒就移动一下，把最新的内容都装进去。这样，系统就能更快地知道用户喜欢什么，推荐也更及时。就像你在玩快节奏的游戏，反应越快，赢的机会越大！
Glossary
Sliding Window (滑动窗口)
一种连续移动的时间段，用于实时采集和处理数据，确保数据的时效性和完整性。技术上实现为连续滑动的时间段，用于平衡时效性和数据完整性。
本文中用来生成实时训练样本，确保标签的时效性。
Re-reco Strategy (重推策略)
一种动态请求推荐服务的机制，周期性地重新请求以减少请求与曝光之间的延迟，提升推荐的时效性。
用于解决请求延迟导致的时效性问题。
Multi-task Learning (多任务学习)
一种同时优化多个相关任务的机器学习方法，通过共享特征或模型参数，提高整体性能。
本文采用多任务模型预测用户多种行为。
CTR (Click-Through Rate, 点击率)
用户点击推荐内容的次数与曝光总次数的比值，用于衡量推荐效果。
作为主要性能指标之一。
NFN (New Follow Number, 新关注数)
在一定时间内新增关注用户的数量，反映内容吸引力。
衡量推荐系统对用户增长的影响。
Fixed Window (固定窗口)
传统数据流处理方式，在预定义的时间段内采集和处理数据。
与本文提出的滑动窗口方案对比。
Latency (延迟)
从用户行为发生到模型响应或推荐内容展示的时间差。
影响推荐系统的实时性。
Label Accuracy (标签准确性)
标签反映真实用户行为的程度，是训练模型的基础指标。
本文关注缩短窗口以提升标签的代表性。
Content CTR (内容点击率)
内容被点击的概率，是衡量内容吸引力的重要指标。
模型预测目标之一。
Multi-modal Transformer (多模态变换器)
结合多种模态信息（如图像、文本、声音）进行特征提取的深度学习模型。
用于提取直播内容的多模态信息。
Open Questions Unanswered questions from this research
- 1 当前滑动窗口机制主要在单一平台验证，跨平台适应性和多场景应用仍需深入研究。未来应结合多模态内容理解和强化学习，提升系统的智能化水平。
- 2 如何自动调节窗口长度以适应不同内容变化速率，是未来的重要研究方向。自适应机制能进一步提升模型的鲁棒性和效率。
- 3 在极端内容变化极快的场景下，滑动窗口的参数调优和模型稳定性问题仍未完全解决。需要开发更智能的调度和优化算法。
- 4 多模态信息融合与时效性保障的结合，是提升推荐质量的关键，但相关技术尚处于早期探索阶段。
- 5 大规模实时系统的计算成本和能耗问题，限制了滑动窗口机制的普及。未来需在算法和硬件层面共同优化。
Applications
Immediate Applications
实时直播内容推荐
结合滑动窗口机制，提升直播平台的内容匹配速度和用户体验，适用于快手、抖音等短视频平台。
电商直播促销
实时分析用户行为，动态调整推荐内容，提升转化率和用户粘性。
在线教育直播互动
根据学生实时反馈，快速调整教学内容和互动策略，增强学习效果。
Long-term Vision
全场景智能推荐系统
结合多模态、多任务、多场景的实时数据流设计，打造全场景智能推荐生态，推动行业数字化转型。
自主学习与自适应优化
引入强化学习和自适应机制，实现推荐系统的自主调节和持续优化，适应内容和用户行为的长期变化。
Abstract
Live streaming recommender system is specifically designed to recommend real-time live streaming of interest to users. Due to the dynamic changes of live content, improving the timeliness of the live streaming recommender system is a critical problem. Intuitively, the timeliness of the data determines the upper bound of the timeliness that models can learn. However, none of the previous works addresses the timeliness problem of the live streaming recommender system from the perspective of data stream design. Employing the conventional fixed window data stream paradigm introduces a trade-off dilemma between labeling accuracy and timeliness. In this paper, we propose a new data stream design paradigm, dubbed Sliver, that addresses the timeliness and accuracy problem of labels by reducing the window size and implementing a sliding window correspondingly. Meanwhile, we propose a time-sensitive re-reco strategy reducing the latency between request and impression to improve the timeliness of the recommendation service and features by periodically requesting the recommendation service. To demonstrate the effectiveness of our approach, we conduct offline experiments on a multi-task live streaming dataset with labeling timestamps collected from the Kuaishou live streaming platform. Experimental results demonstrate that Sliver outperforms two fixed-window data streams with varying window sizes across all targets in four typical multi-task recommendation models. Furthermore, we deployed Sliver on the Kuaishou live streaming platform. Results of the online A/B test show a significant improvement in click-through rate (CTR), and new follow number (NFN), further validating the effectiveness of Sliver.
References (20)
Asymptotically Unbiased Estimation for Delayed Feedback Modeling via Label Correction
Yu Chen, Jiaqi Jin, Hui Zhao et al.
Adam: A Method for Stochastic Optimization
Diederik P. Kingma, Jimmy Ba
Progressive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations
Hongyan Tang, Junning Liu, Ming Zhao et al.
Modeling delayed feedback in display advertising
O. Chapelle
TeRec: A Temporal Recommender System Over Tweet Stream
Cheng Chen, Hongzhi Yin, Junjie Yao et al.
A Nonparametric Delayed Feedback Model for Conversion Rate Prediction
Yuya Yoshikawa, Yusaku Imai
An Overview of Multi-Task Learning in Deep Neural Networks
Sebastian Ruder
Wide & Deep Learning for Recommender Systems
Heng-Tze Cheng, L. Koc, Jeremiah Harmsen et al.
TensorFlow: A system for large-scale machine learning
Martín Abadi, P. Barham, Jianmin Chen et al.
Dynamic Matrix Factorization with Priors on Unknown Values
Robin Devooght, N. Kourtellis, Amin Mantrach
Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate
Xiao Ma, Liqin Zhao, Guan Huang et al.
Online Learning under Delayed Feedback
Pooria Joulani, A. György, Csaba Szepesvari
Estimating conversion rate in display advertising from past erformance data
Kuang-chih Lee, Burkay Orten, A. Dasdan et al.
StreamRec: a real-time recommender system
Badrish Chandramouli, Justin J. Levandoski, Ahmed Eldawy et al.
Online learning for recency search ranking using real-time user feedback
Taesup Moon, Lihong Li, Wei Chu et al.
Understanding the difficulty of training deep feedforward neural networks
Xavier Glorot, Yoshua Bengio
Multitask Learning
R. Caruana
Adaptive Mixtures of Local Experts
R. Jacobs, Michael I. Jordan, S. Nowlan et al.
Billion-scale Commodity Embedding for E-commerce Recommendation in Alibaba
Jizhe Wang, Pipei Huang, Huan Zhao et al.
Streaming CTR Prediction: Rethinking Recommendation Task for Real-World Streaming Data
Qiwen Wang, Hongyu Lu, Yu Chen et al.
Cited By (14)
A Bilateral Perspective for Modeling Real-Time Traffic Trends in Live-Streaming Recommendation
TAGR: Temporally Adaptive Generative Recommendation for Industrial Live-Streaming Advertising
Discovering and Alleviating Data Leakage in Staytime Prediction for Live Streaming Recommendation
Multi-Objective Ranking for Live-Streaming: Balancing Fresh and Delayed Signals with Segment-Aware Targeting
KuaiLive-M3: A Multi-Modal, Multi-Domain, and Multi-Feedback Dataset for Live Streaming Recommendation
SSRLive: Live Streaming Recommendation with Dynamic Semantic ID
Room Matters: Dynamic Room-level Collaboration Information Modeling for Live Streaming Recommendation
Towards Unbiased and Real-Time Staytime Prediction for Live Streaming Recommendation
GRADE: Personalized Multi-Task Fusion via Group-Relative Reinforcement Learning with Adaptive Dirichlet Exploration
KuaiLive: A Real-time Interactive Dataset for Live Streaming Recommendation
LLM-Alignment Live-Streaming Recommendation
CreAgent: Towards Long-Term Evaluation of Recommender System under Platform-Creator Information Asymmetry
Moment&Cross: Next-Generation Real-Time Cross-Domain CTR Prediction for Live-Streaming Recommendation at Kuaishou
MMBee: Live Streaming Gift-Sending Recommendations via Multi-Modal Fusion and Behaviour Expansion