摘要
针对短视频推荐场景中存在的低延迟响应与推荐精度失衡的技术挑战,该研究设计一种双通道动态兴趣建模框架。通过构建基于异构知识图谱的稳态兴趣建模与实时行为感知的瞬态兴趣捕捉协同机制,在确保推荐系统端到端响应时间低于200 ms的工程约束下,有效提升推荐准确度与覆盖率。具体而言,系统采用分阶段兴趣建模策略。在离线阶段,通过3层残差图卷积网络挖掘用户历史交互中隐含的跨实体关联(用户—创作者—主题标签的三元组关系),生成具有语义鲁棒性的长期偏好表征;在在线阶段,部署轻量化流式处理引擎,采用滑动时间窗策略(窗口尺寸动态调整范围为30~120 s)捕捉用户当前会话内的细粒度交互信号(包括视频完播率、互动频率、页面停留时长等12维时序特征),通过门控注意力网络实现短期兴趣的增量式更新。
In response to the technical challenges of low latency response and imbalanced recommendation accuracy in short video recommendation scenarios,this study proposes a dual channel dynamic interest modeling framework.By constructing a transient interest capture collaborative mechanism based on heterogeneous knowledge graph for steady-state interest modeling and real-time behavior perception,the recommendation accuracy and coverage have been effectively improved under the engineering constraint of ensuring that the end-to-end response time of the recommendation system is less than 200ms.Specifically,the system adopts a staged interest modeling strategy:in the offline stage,a three-layer residual graph convolutional network is used to mine the cross entity associations hidden in user historical interactions(user creator topic tag triplet relationship),generating long-term biased representations with semantic robustness;In the online phase,a lightweight streaming processing engine is deployed,using a sliding time window strategy(with a dynamic window size adjustment range of 30~120 seconds)to capture fine-grained interaction signals within the user's current session(including 12 dimensional temporal features such as video completion rate,interaction frequency,and page dwell time),and achieving incremental updates of short-term interests through a gated attention network.
出处
《科技创新与应用》
2025年第10期126-129,共4页
Technology Innovation and Application
关键词
深度学习
短视频推荐
短期兴趣
长期兴趣
双通道动态兴趣建模
deep learning
short video recommendation
short term interests
long term interest
dual channel dynamic interest modeling
作者简介
吉慧(1985-),女,工程硕士,高级讲师。研究方向为计算机数字媒体技术。
