面向公共安全的时空数据挖掘综述

随着各种手持无线设备及传感器的普及,大量的具有时间和空间属性的轨迹数据在不间断地产生.这些不同来源的轨迹数据记录了个体在时间和空间上的活动,从微观和宏观揭示出个人和团体的活动规律,对研究人群行为及城市管理,特别是城市公共安全管理方面,具有重要的意义.以公共安全管理为主要目标,分4个方面调研了相关的研究工作,并分别给出了笔者的研究进展.使用了2类比较有代表性的数据,第1类是智能手机的时间、空间轨迹数据;第2类是城市公共交通卡的换乘数据.第1类是从“点”上分析挖掘个体或者群体的活动规律,而第2类数据则是从“线”上发现人群的聚散规律.基于第1类数据,针对“个体的发现”介绍了相关工作;对于第2类数据,分别从短时和突发2个方面,发现具有潜在危害性的事件,从而向有关部门提供预测和预警,防范该区域可能出现的公共安全事件.比较了各类模型包括经典的时序数学模型ARIMA(autoregressive integrated moving average model)和SARIMA(seasonal autoregressive integrated moving average)、机器学习和神经网络模型SVR(support vector re-gression)、NN(neural networks)、和LSTM(long short-term memorg),发现笔者的模型在短时客流预测方面可以最多提高27.78％,突发客流预测精度可以最高提高到14.68倍.

A survey of data mining on spatial-temporal user behavior data for public safety

With the popularity of smart phones and wireless sensors,large amount of data with timestamps and geo-locations (spatial-temporal) has been produced continuously.This spatial-temporal data records individual behaviors by time and locations,shows macro and micro behavior patterns of people by statistical methods,which is very important for studying the human behavior,especially significant for managing the public safety for city administrators.In this paper,we survey the state-of-the-art research of the human behavior mining for public safety on spatial-temporal data in four aspects,and provide our work in each aspect respectively.We discussed two types of spatial-temporal data,one is smartphone data,and the other is smart card data of public transit.The former shows the individual and crowd behavior from "point" view,and the latter shows the crowd behavior pattern from "line" view.With the former data,we discussed how to discover suspect individuals;with the latter data,we introduced how to find harmful events from short-term and burst passenger traffic,so as to provide the early warning to administration if necessary.We compared our model with existing ones such as ARIMA,SARI-MA,SVR,NN,and LSTM.The result shows that our model can reduce the error up to 27.78％ for short-term traffic prediction,and up to 14.68x for burst traffic prediction.