UNDERSTANDING TAXI SERVICE STRATEGIES FROM TAXI GPS TRACES
UNDERSTANDING TAXI SERVICE STRATEGIES FROM TAXI GPS TRACES
Mining taxi GPS traces has received increasing attentionfrom the data mining, intelligent transportation, database, andubiquitous computing communities [2], [34]. A variety of re-search issues have been addressed leveraging large-scale GPStraces, such as urban human mobility understanding [1], [4],[8], [11], [17], [19], [26], urban planning [16], [29], [30], [35],traffic prediction [9], [23], [24], anomalous trajectory detection[6], [33], city region function identification [20], [22], [28], andtaxi/passenger search strategies [4], [13], [17], [21], [26]. In thissection, we briefly present the related work on improving taxidrivers’ performance, which can be roughly classified into thefollowing three categories.The first category of research focuses on identifying andrecommending popular pickup areas [4], [13], [17], [21], [26].Since the change of taxi status from vacant to occupied isrecorded in the traces and indicates a passenger pickup event,the collection of such events from a large taxi fleet reflects thetaxi demand in the city. By clustering the GPS locations of thesepickup events usingK-meansat different times of the day and indifferent days of the week. Leeet al.[13] analyzed the pickuppatterns of taxis in Jeju, Korea, and recommended the popularclusters to vacant taxis to reduce the idling time of taxis.Changet al.[4] gathered the demand requests according tothe time, location, and weather context; clustered these requestsinto hotspots byK-means,agglomerative hierarchical cluster-ing, andDBSCAN; and recommended the top-ranked placesto taxi drivers. Liet al.[17] predicted the number of pickupevents in different hotspots based on the historical informationrecorded in the taxi GPS traces and provided guidance to vacanttaxis’ drivers.In addition to taking into account the pickup hotspots, thesecond category of research work also considers other prac-tical influencing factors, such as human mobility patterns,passenger-searching possibilities, and potential trip lengths,and provides optimized passenger-searching areas and routes.Unlike the methods focusing on hotspots globally, Powellet al.[21] examined only the surrounding areas. They measured theprofitability of each area in terms of fare gains of all occupiedtrips originated from that area, the number of trips, and the costfrom the current location to that area, exploiting the knowl-edge of passenger’s mobility patterns and taxi drivers’ pickup/drop-off behaviors inferred from taxi GPS traces. Yuanet al.[31]used the historical probability of searching a passenger along aroute to provide drivers with location/route suggestions.Instead of providing explicit guidance about areas or routesfor searching passengers, the last category of research tries toextract effective taxi service guidelines in a city. Velosoet al.[26] investigated the passenger-delivery patterns and passenger-searching processes and revealed that in Lisbon, Portugal, agood passenger-searching strategy in urban areas was that taxisnormally went to adjacent locations, whereas in suburban areas,taxis went to distant locations. By analyzing the taxi GPStraces, Liuet al.[18] uncovered that, in the city of Shenzhen,China, high-revenue taxi drivers had good skills of choosingthe right areas to serve in the city at different times of the dayand selecting the delivery routes with less traffic congestion
These studies are closer to our work than the first two categoriesbecause they provide strategic service guidelines rather thanconcrete areas/routes to taxi drivers. However, [26] did nottackle the passenger-delivery and area-preference strategies,whereas [18] failed to differentiate taxi GPS traces for thepaired drivers. Neither [18] nor [26] built the model charac-terizing the relationship between taxi service strategies and theresulted performance. In addition to leveraging taxi GPS traces,Takayamaet al.[25] used survey information from taxi driversto recommend promising waiting/cruising locations, andYamamotoet al.[27] suggested routing strategies for multipletaxis by mutual exchange of their pathways. However, bothmethods require human intervention.Different from all previous work, we first identify a seriesof taxi service strategies for bothpassenger-searching andpassenger-deliverystrategies. Then, we conduct a correlationanalysis study to extract theefficient and inefficienttaxi servicestrategies at different times and locations. Finally, we build amodel based on L1-SVM using top-ranked service strategiesand validate the effectiveness of the extracted service strategies. click here to view full site
https://goo.gl/maps/1rNVoifRFQ9snSor6
Mining taxi GPS traces has received increasing attentionfrom the data mining, intelligent transportation, database, andubiquitous computing communities [2], [34]. A variety of re-search issues have been addressed leveraging large-scale GPStraces, such as urban human mobility understanding [1], [4],[8], [11], [17], [19], [26], urban planning [16], [29], [30], [35],traffic prediction [9], [23], [24], anomalous trajectory detection[6], [33], city region function identification [20], [22], [28], andtaxi/passenger search strategies [4], [13], [17], [21], [26]. In thissection, we briefly present the related work on improving taxidrivers’ performance, which can be roughly classified into thefollowing three categories.The first category of research focuses on identifying andrecommending popular pickup areas [4], [13], [17], [21], [26].Since the change of taxi status from vacant to occupied isrecorded in the traces and indicates a passenger pickup event,the collection of such events from a large taxi fleet reflects thetaxi demand in the city. By clustering the GPS locations of thesepickup events usingK-meansat different times of the day and indifferent days of the week. Leeet al.[13] analyzed the pickuppatterns of taxis in Jeju, Korea, and recommended the popularclusters to vacant taxis to reduce the idling time of taxis.Changet al.[4] gathered the demand requests according tothe time, location, and weather context; clustered these requestsinto hotspots byK-means,agglomerative hierarchical cluster-ing, andDBSCAN; and recommended the top-ranked placesto taxi drivers. Liet al.[17] predicted the number of pickupevents in different hotspots based on the historical informationrecorded in the taxi GPS traces and provided guidance to vacanttaxis’ drivers.In addition to taking into account the pickup hotspots, thesecond category of research work also considers other prac-tical influencing factors, such as human mobility patterns,passenger-searching possibilities, and potential trip lengths,and provides optimized passenger-searching areas and routes.Unlike the methods focusing on hotspots globally, Powellet al.[21] examined only the surrounding areas. They measured theprofitability of each area in terms of fare gains of all occupiedtrips originated from that area, the number of trips, and the costfrom the current location to that area, exploiting the knowl-edge of passenger’s mobility patterns and taxi drivers’ pickup/drop-off behaviors inferred from taxi GPS traces. Yuanet al.[31]used the historical probability of searching a passenger along aroute to provide drivers with location/route suggestions.Instead of providing explicit guidance about areas or routesfor searching passengers, the last category of research tries toextract effective taxi service guidelines in a city. Velosoet al.[26] investigated the passenger-delivery patterns and passenger-searching processes and revealed that in Lisbon, Portugal, agood passenger-searching strategy in urban areas was that taxisnormally went to adjacent locations, whereas in suburban areas,taxis went to distant locations. By analyzing the taxi GPStraces, Liuet al.[18] uncovered that, in the city of Shenzhen,China, high-revenue taxi drivers had good skills of choosingthe right areas to serve in the city at different times of the dayand selecting the delivery routes with less traffic congestion
https://goo.gl/maps/1rNVoifRFQ9snSor6
Comments
Post a Comment