| 地理位置和时间感知的表示学习框架 |
| |
| 引用本文： | 周俊,,张志强,曹月恬,郑小林.地理位置和时间感知的表示学习框架[J].智能系统学报,2021,16(5):909-917. |
| |
| 作者姓名： | 周俊 张志强 曹月恬 郑小林 |
| |
| 作者单位： | 1. 浙江大学 计算机科学与技术学院， 浙江 杭州 310007;2. 蚂蚁集团， 浙江 杭州 310013 |
| |
| 摘 要： | 现有时空感知的表示学习框架无法对强时空语义的实际场景存在的“When”、“Where”和“What”3个问题给出一个统一的解决方案。同时，现有的时间和空间建模上的研究方案也存在着一定的缺陷，无法在复杂的实际场景中取得最优的性能。为了解决这些问题，本文提出了一个统一的用户表示框架—GTRL (geography and time aware representation learning)，可以同时在时间和空间的维度上对用户的历史行为轨迹进行联合建模。在时间建模上，GTRL采用函数式的时间编码以及连续时间和上下文感知的图注意力网络，在动态的用户行为图上灵活地捕获高阶的结构化时序信息。在空间建模上，GTRL采用了层级化的地理编码和深度历史轨迹建模模块高效地刻画了用户的地理位置偏好。GTRL设计了统一的联合优化方案，同时在交互预测、交互时间预测以及交互位置3个任务上进行模型学习。最后，本文在公开数据集和工业数据集上设计了大量的实验，分别验证了GTRL相较学术界基线模型的优势，以及在实际业务场景中的有效性。
|
| 关 键 词： | 时空语义 时间建模 空间建模 注意力机制 图学习 图神经网络 用户行为建模 用户行为表征 |
A geography and time aware representation learning framework |
| |
| ZHOU Jun,,ZHANG Zhiqiang,CAO Yuetian,ZHENG Xiaoling.A geography and time aware representation learning framework[J].CAAL Transactions on Intelligent Systems,2021,16(5):909-917. |
| |
| Authors: | ZHOU Jun ZHANG Zhiqiang CAO Yuetian ZHENG Xiaoling |
| |
| Affiliation: | 1. College of Computer Science and Technology, Zhejiang University, Hangzhou 310007, China;2. Ant Group, Hangzhou 310013, China |
| |
| Abstract: | The existing geography and time aware representation learning framework fails to give a unified solution to three problems of “When”, “Where” and “What” in real-world scenarios with strong geography and time aware semantics, and the existing research schemes for time and spatial modeling also have certain defects and thus cannot achieve optimal performance in complex actual scenarios. To fill in these gaps, this paper proposes a unified user representation framework—GTRL (geography and time aware representation learning), which is capable of jointly modeling users’ historical behaviors in both time and spatial dimensions. In terms of time modeling, GTRL adopts a functional time encoding and a continuous time and context aware graph attention network to flexibly exploit high-order and structured temporal information from dynamic user behaviour graphs. In terms of spatial modeling, GTRL uses a hierarchical geography encoding and a deep historical trajectory modelling module to effectively depict users’ preference for geographical location. Moreover, a unified optimization strategy is carefully designed to carry out model learning on three tasks of interaction prediction task, interaction time stamp prediction task and interaction geography prediction. At last, extensive experiments on public and industrial data sets are carried out, demonstrating the advantages of GTUL over the academic baseline model and its effectiveness in actual business scenarios, respectively. |
| |
| Keywords: | semantic-space-time mode time modelling spatial modelling attention mechanism graph learning graph neural network user behavior modeling user behavior representation |
|
| 点击此处可从《智能系统学报》浏览原始摘要信息 |
|
点击此处可从《智能系统学报》下载全文 |
|
