Strategies From Taxi GPS Traces
Strategies From Taxi GPS Traces
Passenger-searching strategies. These strategies con-sider the practical factors affecting a taxi driver’s decisionmaking when searching for new passengers. The taxidriver may prefer to wait at a popular location nearby (e.g.,grand hotels and railway stations) or go to a familiar areafar away to search for new passengers.•Passenger-delivery strategies. These strategies character-ize the taxi drivers’ intelligence in delivering passengers.A driver may choose one route among several possibleones to deliver passengers, considering factors such as thetraffic conditions, and trip fares.•Service area preference. In a particular time period of the day, taxi drivers may prefer to serve in certain regions ina city by considering the traffic, passenger demands, and their familiarity with the regions.In order to link the service strategies to the resulted rev-enue of taxi drivers, we need to address the following fourchallenges.The first issue is to separate the GPS traces of individualtaxi drivers so that each taxi driver’s service behaviors can be investigated and linked to the revenue. In Hangzhou, almostall the taxis are shared by two drivers, and these drivers takeshift handover twice every day (i.e., once in the morningand once in the afternoon), but the exact shift handover timechanges from day to day and is not indicated in the GPSrecords.The second issue is to measure a taxi driver’s performance interms of the revenue per operation time slot and to investigatewhether a taxi driver’s operation performance is consistentacross all the time slots. This issue affects the selection of timegranularity. Specifically, if a driver’s operation performance isconsistent on a daily basis, we simply use the average dailyrevenue as the performance indicator; otherwise, we need toinvestigate the correlations between the service strategies andthe average revenue at each time slot rather than on a daily basis.Sri Siva Sakthi Travels
The third issue is to extract and model the service strategiesof taxi drivers. As the service strategies reflect the key decisionsof taxi drivers at different time and locations, it is not alwayseasy to describe and represent them, particularly when theyconsist of a sequence of decisions. For example, for passenger-searching strategies, a taxi driver may make a sequence ofdecisions after dropping off the current passengers and beforepicking up the next passengers; we only investigate the firstintention of the taxi driver and use it to represent his/herpassenger-searching service strategy.The last issue is to discover the efficient and inefficientservice strategies. In our preliminary study presented in [14],we first built a feature matrix, in which each column representsa service strategy (a feature) and each row a taxi. L1-SVMhighlights the salient strategies that can effectively separategood and ordinary taxis.
This is because L1-SVM can selectstrategies and assign the selected strategies with positive andnegative weights to reflect their contributions in differentiatingthe two taxi groups. This is to say that the weights can betreated as the indicators of whether the corresponding strategiesare efficient (ranked positive) or inefficient (ranked negative).However, as noted by Guyon and Elisseeff [10], a featurethat is completely useless by itself can provide a significantperformance improvement when taken with others. Therefore,those highly weighted strategies may be useless when workingalone. Thus, it might be insufficient to merely use the L1-SVMweights to measure the efficiency of the strategies.The main contributions of this paper can be summarized asfollows.• We propose a mechanism to identify when and where taxidrivers take shift handover in order to separate individual taxi drivers’ GPS trajectories. We further propose an algo-rithm to extract the initial intended location of taxi driversafter each drop-off event.• We investigate the taxi service strategies from threeperspectives, namely, passenger-searching strategies,passenger-delivery strategies, and service-region prefer-ence. In particular, we propose three concrete strategiesfor passenger searching, i.e.,hunting locally, waitinglocally, and going distant.• We uncover the fact that most of taxi drivers do notperform consistently across all the time slots. Thus, thecorrelation between the service strategies and drivers’performance in terms of revenue should be studied at the granularity of each time slot.• We represent the taxi service strategies with a feature matrix and use the correlation between different servicestrategies and their corresponding revenue to reveal which strategies areefficient and inefficient. By predicting therevenue of taxi drivers and achieving a prediction residualof less than 2.35 RMB/h, we show that the extracted taxiservice strategies with our proposed approach can wellcharacterize the driving behavior and performance of taxidrivers
https://srisivasakthitravels.com/
https://goo.gl/maps/1rNVoifRFQ9snSor6
Passenger-searching strategies. These strategies con-sider the practical factors affecting a taxi driver’s decisionmaking when searching for new passengers. The taxidriver may prefer to wait at a popular location nearby (e.g.,grand hotels and railway stations) or go to a familiar areafar away to search for new passengers.•Passenger-delivery strategies. These strategies character-ize the taxi drivers’ intelligence in delivering passengers.A driver may choose one route among several possibleones to deliver passengers, considering factors such as thetraffic conditions, and trip fares.•Service area preference. In a particular time period of the day, taxi drivers may prefer to serve in certain regions ina city by considering the traffic, passenger demands, and their familiarity with the regions.In order to link the service strategies to the resulted rev-enue of taxi drivers, we need to address the following fourchallenges.The first issue is to separate the GPS traces of individualtaxi drivers so that each taxi driver’s service behaviors can be investigated and linked to the revenue. In Hangzhou, almostall the taxis are shared by two drivers, and these drivers takeshift handover twice every day (i.e., once in the morningand once in the afternoon), but the exact shift handover timechanges from day to day and is not indicated in the GPSrecords.The second issue is to measure a taxi driver’s performance interms of the revenue per operation time slot and to investigatewhether a taxi driver’s operation performance is consistentacross all the time slots. This issue affects the selection of timegranularity. Specifically, if a driver’s operation performance isconsistent on a daily basis, we simply use the average dailyrevenue as the performance indicator; otherwise, we need toinvestigate the correlations between the service strategies andthe average revenue at each time slot rather than on a daily basis.Sri Siva Sakthi Travels
The third issue is to extract and model the service strategiesof taxi drivers. As the service strategies reflect the key decisionsof taxi drivers at different time and locations, it is not alwayseasy to describe and represent them, particularly when theyconsist of a sequence of decisions. For example, for passenger-searching strategies, a taxi driver may make a sequence ofdecisions after dropping off the current passengers and beforepicking up the next passengers; we only investigate the firstintention of the taxi driver and use it to represent his/herpassenger-searching service strategy.The last issue is to discover the efficient and inefficientservice strategies. In our preliminary study presented in [14],we first built a feature matrix, in which each column representsa service strategy (a feature) and each row a taxi. L1-SVMhighlights the salient strategies that can effectively separategood and ordinary taxis.
This is because L1-SVM can selectstrategies and assign the selected strategies with positive andnegative weights to reflect their contributions in differentiatingthe two taxi groups. This is to say that the weights can betreated as the indicators of whether the corresponding strategiesare efficient (ranked positive) or inefficient (ranked negative).However, as noted by Guyon and Elisseeff [10], a featurethat is completely useless by itself can provide a significantperformance improvement when taken with others. Therefore,those highly weighted strategies may be useless when workingalone. Thus, it might be insufficient to merely use the L1-SVMweights to measure the efficiency of the strategies.The main contributions of this paper can be summarized asfollows.• We propose a mechanism to identify when and where taxidrivers take shift handover in order to separate individual taxi drivers’ GPS trajectories. We further propose an algo-rithm to extract the initial intended location of taxi driversafter each drop-off event.• We investigate the taxi service strategies from threeperspectives, namely, passenger-searching strategies,passenger-delivery strategies, and service-region prefer-ence. In particular, we propose three concrete strategiesfor passenger searching, i.e.,hunting locally, waitinglocally, and going distant.• We uncover the fact that most of taxi drivers do notperform consistently across all the time slots. Thus, thecorrelation between the service strategies and drivers’performance in terms of revenue should be studied at the granularity of each time slot.• We represent the taxi service strategies with a feature matrix and use the correlation between different servicestrategies and their corresponding revenue to reveal which strategies areefficient and inefficient. By predicting therevenue of taxi drivers and achieving a prediction residualof less than 2.35 RMB/h, we show that the extracted taxiservice strategies with our proposed approach can wellcharacterize the driving behavior and performance of taxidrivers
https://srisivasakthitravels.com/
https://goo.gl/maps/1rNVoifRFQ9snSor6
Comments
Post a Comment