计算机科学与探索 ›› 2021, Vol. 15 ›› Issue (9): 1632-1640.DOI: 10.3778/j.issn.1673-9418.2008095

• 学术研究 • 上一篇    下一篇

多角度语义轨迹相似度计算模型

蔡明昕,孙晶,王斌   

  1. 东北大学 计算机科学与工程学院,沈阳 110169
  • 出版日期:2021-09-01 发布日期:2021-09-06

Multi-aspect Semantic Trajectory Similarity Computation Model

CAI Mingxin, SUN Jing, WANG Bin   

  1. College of Computer Science and Engineering, Northeastern University, Shenyang 110169, China
  • Online:2021-09-01 Published:2021-09-06

摘要:

移动设备的发展使得轨迹数据可以记录更多有用的信息,比如签到信息、活动信息,构成了语义轨迹数据。快速有效的轨迹相似度计算会为分析问题带来巨大好处,已有学者对轨迹相似性及语义轨迹相似性做出研究,并提出了一些有效的方法。但是现有轨迹相似性计算方法无法应用于语义轨迹数据,而目前的语义轨迹相似性计算方法又在轨迹采样频率低的情况下效果不佳。因此在解决轨迹相似性计算对低采样频率敏感的基础上,结合了语义轨迹的附加访问地点信息,提出了一种新的轨迹相似性计算模型,叫作多角度语义轨迹(MAST)相似度计算。模型基于LSTM并且引入自注意力机制,学习到的轨迹表达为多个关注轨迹不同方面的低维向量,构成了一个矩阵,从而解决了单一向量无法准确表达轨迹的问题。这个矩阵不仅包含轨迹的空间信息,也包含语义信息,可用于计算语义轨迹相似度。提出的模型在两个现实语义轨迹数据集上进行实验,实验数据表明MAST的计算结果优于现有方法。

关键词: 轨迹相似度计算, 语义轨迹, 自注意力机制, 深度表示学习

Abstract:

The development of mobile devices enables trajectory data to record more useful information, such as check in information and activity information, constituting semantic trajectory data. Fast and effective trajectory similarity computation will bring great benefits to the analysis of problems. Scholars have studied trajectory similarity and semantic trajectory similarity and have proposed some effective methods. However, existing trajectory similarity computation methods cannot be applied to semantic trajectory data, and the current semantic trajectory similarity computation methods do not work well under the condition of low trajectory sampling frequency. In this paper, on the basis of solving the sensitivity of trajectory similarity computation to low sampling frequency, combined with the additional visited point information of semantic trajectory, a new trajectory similarity computation model is proposed, which is called multi-aspect semantic trajectory (MAST). The model is based on LSTM (long short-term memory) and introduces the self-attention mechanism. The learned trajectory is expressed as multiple low-dimensional vectors of different aspects of the trajectory, forming a matrix, thereby solving the problem that a single vector cannot accurately express the trajectory. This matrix contains not only the spatial information of the trajectory, but also the semantic information, which can be used to calculate the similarity of the semantic trajectory. MAST is tested on two realistic semantic trajectory datasets. Experimental data show that MAST is superior to existing methods.

Key words: trajectory similarity computation, semantic trajectory, self-attention mechanism, deep representation learning