Mining Regular Route from GPS Data for Ridesharing Recommendation

1. Mining Regular Route from GPS Data for Ridesharing Recommendation 2012.09.19

2. Author • Wen He TsinhuaUniversity, Beijing, China and Xi'an Communication Institute, Xi'an, China • DeyiLi TsinhuaUniversity, Beijing, China and Chinese Institute of Electronic System Engineering, Beijing, China • TianleiZhang TsinhuaUniversity, Beijing, China • LifengAn TsinhuaUniversity, Beijing, China • Mu GuoTsinhuaUniversity, Beijing, China • GuishengChenChinese Institute of Electronic System Engineering, Beijing, China

3. Outline • Introduction • Related Work • Architecture • The mining of regular routing • Ridesharing recommendations • Experiment discussion • Conclusion

4. Introduction • Why do this research • Improve traffic problem in Beijing, China. • Current ridesharing is still short supplies due to • Journey reliability. • Ridesharing less desirable • Current situation: More people are likely to record daily trajectory. • Current challenge: Regular Route(RP) is difficult recognized from frequent route.

5. Contribution • Made a ridesharing recommendations according to a group of user’s regular routes(RP). • Propose a method to split the mixed user trajectories into each individual route. • Propose a frequency-based regular routes mining method to infer user’s RPs. • Improve the accuracy in distinguishing travel modes between public transports and private driving • Evaluate a method using a large GPS dataset which is provided by GeoLife. This dataset contains 178 realistic user GPS trajectories over a period of four years.

6. Related Work(1/2) • Past Ridesharing Recommendation • A location-based cab-sharing service to help reduce cab fare costs and effectively utilize available cabs • Improve the quality of ridesharing by increasing the driver’s income • Multiple riders was proposed based on the Bee Colony Optimization Metaheuristicmethod

7. Related Work(2/2) • Mining Route History • not only to driver users, but also to users who take public transport as one of their common travel modes. • Our sharable routes are not directly generated based on one day’s trajectory log. • flexible time interval • it’s difficult to find two cars that are always keep synchronized, even they started at the same time and were running on the same road. • The difference between a personal route and a regular route(RP) is that, a personal route does not consider the time factor, and is not a complete route.

8. Architecture • Three components: • Routes Processing • Regular Routes Mining • Ridesharing Recommendation User-based components only need to be preformed once while a user submitting his/her logs to the system.

9. Routes Processing • Stay Regions Subtracting • A stay regions is definitely not a part of a regular • Grids Mapping • combine the time information with grids, and a series of temporal grids is built. • Routes Splitting • segment the trajectory into each individual route.

10. Regular Routes Mining • Routes Grouping • group the routes which happened at similar times of a day together. • Regular Routes Finding: • a frequency-based regular routes mining algorithm is proposed. • Travel Modes Recognizing • A feature of fixed stop rate (FSR) is used to recognize the different travel modes of an RR

11. Ridesharing Recommendations • Grid-based Routes Table Building • For a regular route generated by public transport, we only record it at the starting and ending grids. • Routes Matching With the grid-based • Search two routes which appeared in pairs and also have similar time properties.

12. THE MINING OF REGULAR ROUTES • Routes Processing • two cases that a sequence of GPS points should be split: • user may arrive at his destination, and when he left, a new route will begin. • GPS device was shut down or lost satellite signal over a certain time.

13. Routes Processing Steps(1/3) • Stay Region Subtracting • Pm is just the ending point of the route which enters into a stay region • Pn+1 is the starting point when user departs from the stay region. • we do not denote this region by a single point, but by a pair of indicators (Pm ,Pn), where Pm and Pn are the beginning and ending points of the stay region.

14. Routes Processing Steps(2/3) • Grids Mapping

15. Routes Processing Steps(3/3) • Routes Splitting

16. Regular Routes Mining(1/2) • An RR is a complete route where a user frequently passed through in approximately the same time of day. • The first is , which is used to decide the frequency of a route, • and the second is , which is used to decide a similar time.

17. Regular Routes Mining(2/2) • Routes Grouping • Therefore it's difficult to extract RRs from all routes directly • But an RR should always happen at a similar time of day. • group routes not only based on the time of day but also the day of the week

18. Regular routes Finding(1/7) • After grids mapping, the trajectories are formed as R1, R2 and R3 in Figure4 (b).

19. Regular routes Finding(2/7) • We say a DE is a FDE if DE.num is larger than threshold of . • RR is a route which is frequently visited by a couple of complete routes, but not some parts of a route. This means we should not directly use FDEs to represent an RR

20. Regular routes Finding(3/7) • In a set of t-Routes, FDEs may exist without an RR. • But if there is an RR, the RR will have large common parts with FDE

21. Regular routes Finding(4/7) • A frequency-based regular route mining method • 1. Calculate frequent coefficient (FC) of each route • The frequent coefficient is defined as fc(R) =m/n, where n is the number of DEs in the route R, m is the number of FDEs in the route R. • 2. Find frequent routes • A route with fc(R)> fcthreh will be deemed as a frequent route. • 3. Calculate regular coefficient (RC) of each FDE

22. Regular routes Finding(5/7) • A frequency-based regular route mining method • 4. Find Regular FDEs (RFDE) • 5. Use RFDEs instead of FDEs to repeat step 2 to 4.

23. Regular routes Finding(6/7) • in Figure5 (a), both R2 and R4 passed the DE (JS->JT), but since R4 is not a frequent route, it has no contribution to an RR, DE(JS->JT) cannot be an RFDE.

24. Regular routes Finding(7/7) • Add time property (ts, td) for each RR, where ts and td denote the start and the duration time of the route respectively where n is the number of the support routes of an RR.

25. Travel Modes Recogning(1/3) • According to make a recommendation for ridesharing, there are two transport modes: • public transport • private driving.

26. Travel Modes Recogning(2/3) • distinguish different transport modes • Which one is public transport • frequently at fixed regions like bus stops or subway stations • Then an RFDE with SP lower than is a stop region.

27. Travel Modes Recogning(3/3) • Fixed Stop rate(FSR) • the number of stop regions within a certain distance

28. RIDESHARING RECOMMENDATIONS • Grid-based Routes Table Building • if it is generatedby public transportation, it will only be recorded in its origin anddestination grids.

29. RIDESHARING RECOMMENDATIONS • Routes Matching • Two kind of car sharing • Public transportation • Private driving • if a query route is generated by public transport, only routes by driving modes could be recommended

30. RIDESHARING RECOMMENDATIONS • Flow chart if the process

31. EXPERIMENTS DISCUSSION • This dataset is consisted of 178 users' realistic trips over a period of 4 years (from 2007 to 2011). • Most of the time, we see all mined RRs as different users’ RRs

32. Experiment Result • Influence of grid size • Smaller the grid size, the larger the storage space is needed • Too larger a grid size will lost some details • 10 sec as final grid size in our experiment

33. Experiment Result • 3 routes are support routes of RR from 9 routes in similar time. • Robust to slight disturbance Compare with ANTrip trajectory

34. Experiment Result • the trajectories are generated by bus • Only two one RR for the user (according to three bus routes(a))

35. Experiment Result • are too short to make a ridesharing • RR are dense in north of Beijing in Microsoft Research

36. Experiment Result • using FSR to distinguish traffic modes between public transportation and driving. • the accuracy could reach 0.876

37. Experiment Result • Routes matching. • (a) and (c) are public transportation • (b) and (d) are driving

38. Experiment Result • The storage requirement of the proposed method. • The first row is the number of records • The second row is the storage ration between the numbers of the original dataset. • The storage requirement is quite lightweight

39. Conclusion • A frequency-based regular route mining algorithm is proposed • each part of a regular route must be visited frequently. • a regular route should be frequently visited by some complete routes called support routes. • most parts of a support route must pass through the frequently visited regions • identified to distinguish travel modes between public transportation and individual driving • Valuated on a real-world GPS dataset, which is consisted of 178 users over a period of 4 years

40. Futrue work • More flexible ridesharing strategies will be considered in our future work. • find a route which reaches at his/her nearest subway station.

41. BACKUP Presented by Ivan Chiou