1. C ROWD P LANNER : A C ROWD -B ASED R OUTE R ECOMMENDATION S YSTEM Han Su, Kai Zheng, Jiamin Huang, Hoyoung Jeung, Lei Chen, Xiaofang Zhou

2. C ONTENT 1 Introduction 2 Overview of CrowdPlanner 3 Preliminary 4 Task Generation 5 Worker Selection 6 Experiments 7 Conclusion 2

3. I NTRODUCTION With the development of GPS technologies and a number of navigation service providers (e.g., Google Map, Bing Map, Naver): Travel to unfamiliar places with much less effort. Select shortest route and/or fastest routes. Are these routes always good enough to be the best choice when people travel ? There are substantial differences between popular routes and recommended routes by experienced/frequent drivers. Drivers’ preferences are influenced by lots of additional factors: the number of traffic lights, speed limitation, road condition, weather, etc. This paper proposes a novel crowd-based route recommendation system, CrowdPlanner: Take the emerging concept of crowd sourcing. Blend domain-expert knowledge for route recommendation. 3

4. I NTRODUCTION Difficulties of crowd-based route recommendation: How to automatically generate a user-friendly task? => For doing the job more comfortably. How to choose a set of suitable worker for a given task? => For improve the performance of the recommendation. Solution: Utilizes a set of discriminative land-marks to generate a binary question set by analyzing the given candidate route set. Identify a few key attributes of workers that mostly affect their performance on a given task. Propose an efficient search algorithm to find the most eligible workers. 4

5. C ONTENT 1 Introduction 2 Overview of CrowdPlanner 3 Preliminary 4 Task Generation 5 Worker Selection 6 Experiments 7 Conclusion 5

6. O VERVIEW OF C ROWD P LANNER Traditional Route Recommendation Module (TR): Control logic component : trigger Reuse truth and Route Evaluation Receive the user request and control the workflow. Return result immediately if the request is hit of the truth. 6

7. O VERVIEW OF C ROWD P LANNER Traditional Route Recommendation Module (TR): Route generation component : generate candidate routes. Route evaluation component : trigger route generation component Evaluate the routes using computer power and reduce the cost of CrowdPlanner. -> If some of recommended routes agree with each other to a high degree, one of them will be selected as the best recommended route. -> A route with the highest confidence score that is greater than a threshold η will be selected as the best recommended route. Otherwise, trigger CrownPlanner. 7

8. O VERVIEW OF C ROWD P LANNER Crowd Route Recommendation Module (CR): Task generation component : Generate a task by proposing a series of questions for workers to answer. Worker selection component : select a set of eligible workers who are most suitable to answer the questions. Early stop component : control the reply time. Rewarding component : rewards the workers according to their workload. 8

9. C ONTENT 1 Introduction 2 Overview of CrowdPlanner 3 Preliminary 4 Task Generation 5 Worker Selection 6 Experiments 7 Conclusion 9

10. P RELIMINARY Landmark ( l ) : a geographical object in the space, which is stable and independent of the recommended routes. Landmark-based Route R = [l 1, l 2,..., l n ] : a route represented as a finite sequence of landmark. Discriminative landmarks : A landmark set L is called discriminative to a set of landmark-based routes R if for any two routes R 1 and R 2 of R, the joint sets R 1 ∩ L and R 2 ∩ L are different. For example, L 1 = { l 2, l 3, l 4 } is discriminative to R 1 = { l 1, l 2, l 3 } and R 2 = { l 1, l 2, l 4 }, since the joint sets R 1 ∩ L 1 = { l 2, l 3 } and R 2 ∩ L 1 = { l 2, l 4 } are different, but L 2 = { l 1, l 2 } is not discriminative to R 1 and R 2. 10

11. C ONTENT 1 Introduction 2 Overview of CrowdPlanner 3 Preliminary 4 Task Generation 5 Worker Selection 6 Experiments 7 Conclusion 11

12. T ASK GENERATION Landmark Significance: Utilize the popular location-based social network (LBSN) and trajectories of cars to infer the significance of landmarks. Landmark Selection: Selected landmark set L should be discriminative to the candidate routes R. -> The difference between any two routes can be presented. The number of landmark should be as small as possible. 12

13. T ASK GENERATION Landmark Selection: Preparation step : filter out some non-beneficial landmarks, i.e., the ones which cannot discriminate any routes. Expansion step : recursively generates the test landmark set S. Test step : Test to see whether S is discriminative. Select the set S that have maximum significant. 13

14. T ASK GENERATION Landmark Selection: The paper also presents some optimization method for reducing the generation time. Question Ordering: It is not necessary to ask all the questions in most cases. E.g: if a worker indicates that she prefers the routes passing l 2 from l 1 to l 10, we do not need to ask whether he recommend to pass l 8 since all the routes passing l 2 do not pass l 8 Compute and sort the questions based on the informativeness. Arrange the questions into a tree-like structure. 14

15. C ONTENT 1 Introduction 2 Overview of CrowdPlanner 3 Preliminary 4 Task Generation 5 Worker Selection 6 Experiments 7 Conclusion 15

16. W ORKER SELECTION 16 Response time: Each task has a user-specified response time. Probability of a worker to respond a task within time t is defined by If F is less than the threshold η time, we will not assign the task to him. Worker’s Familiarity Score: The accumulated familiarity score:

17. W ORKER SELECTION 17 Sort the candidate workers in descending order based on the familiarity score. The preference score of l j to each w is defined as follow: W lj is the workers who have non-zero accumulated familiar scores. Worker gets the maximum sum of the preferences will be assigned the job.

18. C ONTENT 18 1 Introduction 2 Overview of CrowdPlanner 3 Preliminary 4 Task Generation 5 Worker Selection 6 Experiments 7 Conclusion

19. E XPERIMENTS 19 Experiment Setup: Trajectory Dataset: real trajectory datasets generated by taxis, trucks and private cars in Beijing and Nanjing. POI Clusters: two POI datasets of the Beijing and Nanjing cities. Ground truth route: 1000 popular routes agreed by all three route mining algorithms. Workers: volunteers.

20. E XPERIMENTS 20

21. E XPERIMENTS 21 MO: show the candidate routes directly on map and ask workers to choose. CB: workers need to choose all the landmarks on their preferred routes. BwO: the questions are asked in the descending order of the significance. BO: the proposed question format.

22. E XPERIMENTS 22

23. C ONCLUSION 23 This paper proposed a novel crowd-based route recommendation system – CrowdPlanner. Two core components: task generation and worker selection. CrowdPlanner is able to recommend users the most satisfactory routes with at least 90-percent chances. This research sheds light on some other crowd-based recommendation systems also.

24. 24 THANK YOU