顾及相似用户特征的个人位置预测算法

个体位置预测在传染病精准防控、公共设施科学规划等应用中具有重要意义，既有位置预测算法主要侧重对个体纵向历史轨迹特征进行挖掘建模，从而实现位置预测，对横向相似性用户的规律特性考虑较少。因此，基于图卷积和长短期记忆模型（long short-term memory,LSTM）框架，提出顾及横向相似用户轨迹特征以及纵向历史规律性特征的个体位置预测算法。首先，构建用户轨迹相似性算法并筛选高相似度用户；然后，利用图卷积模型提取待预测用户相似高的用户轨迹特征；最后，利用LSTM框架提取历史轨迹特征，集成相似用户轨迹特征，从而实现个体位置预测。基于某市8万多个用户连续4个工作日的数据进行实验，结果表明，所提算法的准确率随预测时间步长增加而下降，而夜间预测准确率明显高于白天，但相比于既有模型均有10%以上提高；以15 min为预测时间步长时，模型准确率达80.45%。

An Individual Location Prediction Algorithm Considering Similar User Characteristics

  Objectives  Individual location prediction is significant in applications such as precise prevention, control of infectious diseases and scientific planning of public facilities. Existing location prediction algorithms mainly focus on mining and modeling individual longitudinal historical trajectory characteristics, and realize location prediction, and less consider the regular characteristics of users with horizontal similarity. Therefore, based on the framework of graph convolution network and long short-term memory (LSTM) model, this paper proposes an individual location prediction algorithm considering the characteristics of horizontally similar user trajectories and the characteristics of vertical historical regularity.  Methods  First, we construct a user trajectory similarity algorithm, and select screen users with high similarity. Second, we use the graph convolution model to extract user trajectory features with high similarity. Finally, we use the LSTM model framework to extract historical trajectory features and integrate similar user trajectory features, so as to achieve individual location prediction.  Results and Conclusions  Based on the data of more than 80 000 users in a city for 4 consecutive working days, the results show that the accuracy of the proposed algorithm decreases with the increase of the prediction time step, and the accuracy of night prediction is significantly higher than that of the day. Compared to the previous models, our proposed algorithm has an improvement of more than 10%. When 15 minutes is used as the prediction time step, the model accuracy rate reaches 80.45%.