《Hunting or waiting?Discovering passenger finding strategies from a large-scale real-world taxi dataset》2011 IEEE
Abstract:
本文目标是从出租车GPS轨迹数据中发掘有效的和无效的passenger-finding strategies；本文构建了Location-Strategy feature triplet的特征，用于形式代表这些strategies；本文利用L1-Norm SVM 筛选出决定taxi performance的最显著的特征；最后对比 用筛选出的特征预测taxi performance的精度 和 用全部特征所预测的taxi performance的精度，发现 用筛选出的特征预测taxi performance的 性能更好，说明所筛选的特征是准确的。
BASELINE：
1、提出了与上车点和下车点相关的两个distance
1、 the driving distance during 3 minutes before a pickup event （上车事件之前3分钟的驾驶距离）
2、the driving distance during the period after current dropoff event and before the next pickup event（在当前下车事件之后和下一次上车事件之前的驾驶距离）
如图所示：
For example
the first row means that “at 00:19:48, 12 Oct 2009, taxi 20731 picked up a passenger at [120.107590, 30.320194], and 3 minutes before this pickup event, the taxi had run 0.68km”.
The second row means that “at 00:38:15, 12 Oct 2009, taxi 20731 dropped off a passenger at [120.191025, 30.265770], and from this dropoff event to the next pickup event, the taxi had run 1.99km” .
2、基于grid的出租车上下车事件hotmap
3、构建了Time-Location-Strategy feature triplet的Descriptors
(Time, Location, Strategy)
For “Time”, we divide one day into 12 equal intervals and the resulting time slots are {00h~02h, 02h~04h, …, 22h~24h}. For “Location”, we use the region labels {1, …, 100} defined in Figure 1, in which each cell is approximately a 1200×1200 m2 area. For “Strategy”, it is a boolean value for indicating that a taxi driver is“hunting” or “waiting” passengers before a pickup or chooses “local” or “distance” to find the next passenger after a dropoff.
4、基于L1-Norm SVM挖掘Good/Bad Strategy
Based on the obtained training examples and their labels, we use L1-Norm SVM [15] to learn a classifier as well as select a small subset of most salient features for good and ordinary taxis from a collection of the taxi-patterns.
结果如下:
5、文中提到的有一些后续研究价值的文献
[8] Chang, H., Tai, Y., and Hsu, J.Y. “Context-aware taxi demand hotspots prediction,”.Int. J. Bus. Intell. Data Min., Vol.5, No.1, pages 3-18,2010.
[14] Yamamoto, K., Uesugi, K., and Watanabe, T.“Adaptive routing of multiple taxis by mutual exchange of pathways”,Int. J. Knowl. Eng. Soft Data Paradigm, Vol. 2,No. 1, pages 57-69, 2010.
思考:
文中提到的与上车点和下车点相关的两个distance:
1、 the driving distance during 3 minutes before a pickup event （上车事件之前3分钟的驾驶距离）
2、the driving distance during the period after current dropoff event and before the next pickup event（在当前下车事件之后和下一次上车事件之前的驾驶距离）
可以考虑加入到我们的特征构建中