GoogleLocationHistoryDataKey.csv

In order to compare GLH data to travel diary data, GLH data were captured passively across five smartphone devices<b> </b>while two researchers followed prescribed itineraries. While the same five smartphone devices were not all carried on each prescribed itinerary, resulting in an uneven number of samples between devices, two researchers traveled together along prescribed itineraries following identical characteristics of travel. Arrival and departure times from all prescribed itinerary locations were measured empirically by self-report, thus completing the prescribed itineraries as travel diaries. This method afforded <i>a priori </i>knowledge of what the GLH data should be reporting as expected from the prescribed itineraries. In all, 12 unique itineraries were prepared, accounting information that GLH data provide, and to which can be compared: positions of locations, arrival and departure times, location dwell times, trip duration times, trip distances, travel modes, and trip geometries throughout different urban forms. <br> <br> Once GLH data were captured, the downloadable Keyhole Markup Language (KML) files were retrieved from the desktop Google Maps “Your Timeline” feature. The attribute information associated to the point and line shapefiles (SHP) derived from the KML files were then matched to the itinerary data by arrival and departure times within an error of ten minutes. We accepted an error of ten minutes as a reasonable limit for which we could unambiguously assume that the GLH data were sensing our prescribed itinerary travel and assign a match.<br> <br> The following analysis were done for each variable interaction of interest between smartphone device, travel mode and urban form. Mean location position errors were calculated by taking the mean Euclidean distance between the expected locations and the matched GLH data locations. Arrival and departure times were assessed using the root mean square error (RMSE). Similarly, trip duration times, trip distances, and location dwell times were assessed by the normalized RMSE to account for the differences in range for each interaction subset. Finally, GLH data trip geometries were assessed using the mean Hausdorff distance. The Hausdorff distance is a measure of the maximum distance from a point in set A to its nearest point in set B. In other words, of all the shortest distances from each point in set A to points in set B, it is the maximum of those distances. Since GLH data trip geometries were given as linestrings, we converted the linestrings to point sets for computing Hausdorff distances. This was done by generating points along linestring geometries. An interval of one meter between points was chosen as a distance small enough to account for the geometric variation in trip geometry for all scales of travel. The mean Hausdorff distances were then calculated for each interaction subset of travel mode and urban form.

为对比GLH数据与旅行日志数据的一致性，研究团队依托5台智能手机设备被动采集GLH数据，同时由两名研究人员严格遵循预设行程开展出行。由于单次预设行程未全程携带全部5台智能手机，导致不同设备间的样本量存在不均衡情况，但两名研究人员始终结伴按照完全一致的出行特征完成预设行程。所有预设行程途经点位的到达、离开时间均通过自主上报完成实测记录，以此构建旅行日志形式的行程数据。该方法可使研究者先验（a priori）获知GLH数据应按照预设行程上报的预期内容。本次研究共规划12条独立行程，涵盖GLH数据可提供的全部信息维度，用于开展对比分析：各点位位置、到达与离开时间、点位驻留时长、行程时长、行程距离、出行方式，以及不同城市形态下的行程几何特征。

完成GLH数据采集后，研究人员从桌面版谷歌地图（Google Maps）的“你的时间线”功能中导出可下载的地标标记语言（Keyhole Markup Language, KML）文件。随后将KML文件转换得到的点、线矢量形状文件（Shapefile, SHP）所关联的属性信息，以10分钟以内的时间误差为阈值，与预设行程数据进行匹配。我们将10分钟误差阈值设定为合理范围，借此可明确判定GLH数据已感应到预设行程出行，并完成匹配关联。

针对智能手机设备、出行方式与城市形态间的各目标变量交互关系，开展如下分析：通过计算预期点位与匹配后的GLH数据点位之间的平均欧氏距离，得到平均点位位置误差；利用均方根误差（Root Mean Square Error, RMSE）评估到达与离开时间的偏差。类似地，针对行程时长、行程距离以及点位驻留时长，采用归一化均方根误差（Normalized RMSE）进行评估，以消除不同交互子集间数值范围差异带来的影响。最后，针对GLH数据的行程几何特征，采用平均豪斯多夫距离（mean Hausdorff distance）进行评估。豪斯多夫距离（Hausdorff distance）用于衡量集合A中任意一点到集合B中最近点的最大距离，换言之，即集合A中所有点到集合B的最短距离中的最大值。由于GLH数据的行程几何以线串（linestring）形式给出，我们将线串转换为点集以计算豪斯多夫距离：沿线串几何生成间隔为1米的采样点，该间隔足够小以覆盖所有出行尺度下的行程几何变化。随后针对出行方式与城市形态的各交互子集，计算其平均豪斯多夫距离。