TrajectMe - Planning Sightseeing Tours with Hotel Selection from Trajectory Data

In this article, we propose TrajectMe, an algorithm that solves the orienteering problem with hotel selection in several cities, taking advantage of the tourists' trajectories extracted from location-based services. This method is an extension of the state-of-the-art memetic-based algorithm. To this end, we collect data from Foursquare and Flickr location-based services, reconstruct the trajectories of tourists. Next, we build a hotel graph model (HGM) using a set of trajectories and a set of hotels to infer typical sequences of hotels and point of interest (PoI). The HGM is applied in the initialization phase and in the genetic operations of the memetic algorithm to provide good sequences of hotels, whereas the associated sequence of PoIs are improved by applying local search moves. We evaluate our proposal using a large and real dataset from three Italian cities using up to 1000 hotels. The results show that our approach is effective and outperforms the state-of-the-art when using large real datasets. Our approach is better than the baseline algorithm by up to 208% concerning the solution score and proved to be more profitable toward PoI visiting time, being 54% better than state-of-the-art.

Transcript

1. AQUI FICA O TÍTULO DE SUA APRESENTAÇÃO Data Science Laboratory Federal University of Ceará
2. TrajectMe Planning Sightseeing Tours with Hotel Selection from Trajectory Data Emanuel Oliveira, Igo Brilhante and José Macedo
3. 1. Motivation 2. Problem Statement 3. Related Work 4. TrajectMe Algorithm 5. Experiments & Results 6. Conclusion AGENDA
4. 4 Motivation Tuscany, Italy Font: https://commons.wikimedia.org
5. 5 MOTIVATION Tuscany, Italy Font: http://welcometuscany.it
6. 6 MOTIVATION Tuscany, Italy Font: http://welcometuscany.it
7. 7 MOTIVATION Tuscany, Italy Font: http://welcometuscany.it
8. 8 CHALLENGE For each day of this travel, the start and end hotel must be selected optimally • Maximizing the satisfaction and the visiting time to the selected PoIs • Minimizing the travel time among PoIs
9. 9 CHALLENGE Hotels can be selected from all suitable hotels available within the city Hotels Available in Tuscany, Italy Font: https://foursquare.com
10. 10 ORIENTEERING PROBLEM WITH HOTEL SELECTION Memetic Algorithm Orienteering Problem with Hotel Selection Orienteering Problem Travelling Salesman Problem Knapsack Problem Team Orienteering Problem (Team) Orienteering Problem with Time Window Skewed Variable Neighborhood Search
11. 11 CHALLENGE ▹ Searching hotels in Rome will return more than 1000 hotel options ¹; ▹ Current OPHS state-of-the-art solutions do not scale on real applications, with hundreds of hotels and PoIs; ▹ Current solutions are limited to compute up to 15 hotels and 100 PoIs; ▹ We need a solution that scales given a large number of hotels. ¹ http://booking.com
12. 12 INTUITION - USE TRAJECTORY DATA
13. 13 CONTRIBUTIONS 1. A solution to the OPHS problem using historical trajectories of tourists (Hotel Graph Model); 2. An efficient algorithm for coping with thousands of hotels; 3. A set of experiments using a real dataset demonstrating the accuracy and efficiency of our approach.
14. 14 PROBLEM STATEMENT A SET OF HOTELS... hotel departure hotel arrival hotel
15. 15 PROBLEM STATEMENT AND A SET OF POIS. hotel departure hotel arrival hotel point of interest
16. 16 PROBLEM STATEMENT SCORE AND TRAVEL TIME hotel departure hotel arrival hotel point of interest Each PoI is associated with a score and a visiting time. Hotels do not have a score The travel time between the each pair of vertex is known.
17. 17 PROBLEM STATEMENT TRIP • Start hotel • Ordered Sequence of PoIs • End Hotel • Limited by a daily budget
18. 18 PROBLEM STATEMENT trip 1 trip 2 trip 3 A SOLUTION • Connected trips • Objective: maximize the score • The final score is the sum of PoIs’ scores.
19. 19 RELATED WORKS TrajectMe Memetic Algorithm Orienteering Problem with Hotel Selection TripBuilder Tourist Trip Design Problem Orienteering Problem Travelling Salesman Problem Knapsack Problem Team Orienteering Problem (Team) Orienteering Problem with Time Window Skewed Variable Neighborhood Search
20. DIVSALAR et al. Skewed Variable Neighborhood Search (SVNS) (2013) • Introduced the OPHS. • Presents mathematical formulations. Memetic Algorithm (MA) (2014) • Two level solution: hotel selection and PoIs selection. • Faster than SVNS in larger instances. BRILHANTE et al. TripBuilder (2015) • Mines data from location-based. • Modeled as instance of the Generalized Maximum Coverage (GMC) problem. • Extract trajectories made by tourists in the past to create routes. • Don’t includes hotel selection as part of planning. 20 RELATED WORKS
21. RELATED WORKS MEMETIC ALGORTITH TRIPBUILDER TRAJECTME 21
22. 22 RELATED WORKS Work Hotels Trajectories Real Datasets Skewed Variable Neighborhood Search (2013) X Memetic Algorithm (2014) X TripBuilder (2015) X X TrajectMe (2018) X X X
23. 23 BUILDING KNOWLEDGE BASE Data collection Data processing Hotel Graph Model Users Photos PoIs and Hotels User PoI History Traject ory Set Hotels Set Building Hotel Graph Model
24. 24 SOLUTION Main Loop Genetic Operators and Local Search. Initialization Generate initial population using the HGM. Hotel Graph Model Highlight the best hotels in the city to generate the trips for OPHS.
25. 1 2 3 STEPS 25 1 hotel point of interest trajectory end (PoI) trajectory HOTEL GRAPH MODEL (HGM)
26. 1 2 3 STEPS 26 1 HOTEL GRAPH MODEL (HGM)
27. 1 2 3 STEPS 27 1 HOTEL GRAPH MODEL (HGM) knn with k equals 1, i.e., 1nn.
28. HOTEL GRAPH MODEL (HGM) 1 2 3 STEPS 28 1
29. 1 2 3 INITIALIZATION - GENERATE INITIAL POPULATION STEPS 29 The creation of initial population can be obtained through possible paths, i.e, sequences of hotels, in HGM. Generate a sequence of hotels Reached the size of population? No Yes Select from departure hotel Select a next hotel Reached the arrival hotel? Yes No The choice of the next hotel favors edges with a higher score.
30. 1 2 3 MAIN LOOP STEPS 30 New solutions is formed by crossovers and mutation. Put offspring to the Pool. 2. Apply Genetic Operators Copy the current population to a Pool of solutions 1. Populate the Pool The best solutions are selected to create the next population, 3. Population Manager Loop until reach the max. of iterations (parameter)
31. 1 2 3 MAIN LOOP - APPLY GENETIC OPERATORS STEPS 31 Select parent solution(s) from population Apply genetic operator Put offspring solution(s) to Pool The choice of the parent solution favors those with a highest scores. • 3 genetic operators ◦ 2 crossovers ◦ 1 mutation • Applied several times in the main loop in such way that the Pool of solutions reach the double size of population. Apply Local Search to improve solution.
32. 1 2 3 MAIN LOOP - APPLY GENETIC OPERATORS STEPS 32 CROSSOVERS MUTATION Parent solution A Parent solution B Crossover Operator Offspring solution A Offspring solution B Parent solution Mutation Operator Offspring solution
33. 1 2 3 MAIN LOOP - MANAGE POPULATION STEPS 33 Solution Solution Solution Solution Solution Solution POOL ORDERED POOL NEXT POPULATION Sorted by solution score. High score solutions has preference, but low score solutions are also included to increase the variation and escape from local optimum.
34. 1 2 3 MAIN LOOP - BEST SOLUTION STEPS 34 For each iteration, the solution with highest score solution is save; At the end of all iterations, the best solution is returned as the result.
35. 35 EXPERIMENTS - DATASETS City PoIs Hotels Trajectories Pisa, Italy 61 402 59 Florence, Italy 146 1000 593 Rome, Italy 302 1000 1685
36. 36 EXPERIMENTS - TOUR SCORE METRIC Evaluates the total score of the PoIs in the solutions. It is the ratio of the total score of the PoIs in the solution tour over the sum of the scores of all the PoIs. Sum of scores of the PoIs in the Tour Sum of all PoIs' scores in the city Tour Score:
37. 37 EXPERIMENTS - TOUR UTILITY METRIC Evaluates how good is the solution in terms of visiting time. Computed as the sum of the visiting time for the PoIs for each trip in the tour divided by available time budget of the day. Higher scored tours are preferable since they favor the time to enjoy attractions with respect to the traveling time. Sum of the visiting time for the PoIs for each trip in the tour Time budget of the day Tour Utility:
38. 38 EXPERIMENTS - IMPROVEMENT METRIC In addition to the computed scores for the evaluation metrics, we present the improvements of TrajectMe over MA.
39. 39 EXPERIMENTS - METHODOLOGY • TM-k is the variant of TrajectMe for k; • We studied the algorithms by varying the number of days and the parameter k for each city; • All results are the averages of three executions; • The best result is highlighted with bold; • We present the score and the improvement of TrajectMe with respect to the MA result;
40. 40 EXPERIMENTS - PISA DATASET • Pisa where only considered 2 days because it is a small city where tourists can visit almost all the PoIs in one or two days. • Results for Pisa show that both algorithms achieved good results, where MA was slightly better than TrajectMe (TM). • This result highlights that for cities with a small set of PoIs, both algorithms tend to perform well since the tasks of selecting hotels and PoIs are simple in this scenario.
41. 41 EXPERIMENTS - FLORENCE DATASET • TM presented the highest scores, in special for 2-days tours, for both cities. • Number of PoIs and hotels are much larger than in Pisa • Finding good tours become a harder task to achieve. • We notice that as the number of days increases, the improvements of TM w.r.t. MA decreases.
42. 42 EXPERIMENTS - ROME DATASET • Large number of PoIs and hotels. • The improvements are still significant. • The MA seem to very affected for the initial population and the number of iterations in the main loop. Require much more iterations to possibly overcome bad initial solutions. • Starting with good population already gives good solutions and few additional iteration are necessary to refine the best solution.
43. 208% of improvement in Tour Score for Rome in 2 days over MA. 43 EXPERIMENTS - RESULTS
44. 54% of improvement in Tour Utility for the same scenario . 44 EXPERIMENTS - RESULTS
45. 48 CONCLUSION • We present TrajectMe, a memetic algorithm boosted with the Hotel Graph Model to solve the orienteering problem with hotel selection (OPHS); • Hotel Graph Model allowed to generate tours for the OPHS more effectively; • We perform experiments using real datasets provided by location-based services including thousands of hotels. • TrajectMe significantly overcomes the competitor achieving better efficiency, mainly when the number of hotels and PoIs are large.
46. 49 FUTURE WORK • Extend our proposal towards a personalized tour generation with hotel selection based on the preferences of the users and categories of the PoIs; • Studying the choice of k and the other parameters used in our experiments to better understand the impact in the results; • Investigate the scenario when not enough trajectory data is available.
47. 47 Thank you! Questions? You can contact me ▹ jose.macedo@dc.ufc.br ▹ igobrilhante@lia.ufc.br ▹ emanueloliveira@lia.ufc.br