«Automated Tourist Decision Support Wouter Souﬀriau Dissertation presented in partial fulﬁllment of the requirements for the degree of Doctor in ...»
The City Trip Planner tourist decision support system suggests personal trips by ﬁrst capturing the tourist’s trip constraints and interests through a small questionnaire. These interests are used to predict a personal interest score for each POI that a tourist can visit within the constraints. The GPS coordinates and opening hours of the POIs, the personal interest scores, the tourist’s location and available time give rise to a speciﬁc instance of a TTDP. Next, a heuristic algorithm solves this instance, resulting in a personal trip, that is tailored to the user’s interests, current location, destination, available time and the opening hours.
Finally, the trajectory can be printed or downloaded to a GPS navigation device.
Figure 6.2 overviews the diﬀerent components of the system.
Figure 6.2: The City Trip Planner - System Overview
1 describes the structure of the tourist POI database, Section 6.2.2 describes how to synthesise personal tourist information into the user proﬁle. Next, Section 6.2.
3 matches this user proﬁle to the pre-determined POI database in order to predict personal interest scores for the diﬀerent POIs, making the formulation of a tourist trip design problem instance possible. The algorithm presented in Section 6.2.4 solves this instance and proposes a trip to the user. When the tourist is happy with the proposal, the trip can be ﬁnalised, as is explained in Section 6.2.5.
6.2.1 POI Database
109. Every POI is characterised by its GPS coordinates, a calendar of opening and closing hours, an average visiting duration and textual descriptions in English and Dutch. A POI belongs to exactly one ”type”: abbeys, beguinages, castles, churches, musea, etc. Each city deﬁned at most ten POI types that are very relevant for the city concerned. Moreover, the POIs have been classiﬁed manually using the following categories: archaeology, architecture, classical art, “local markets and streets”, modern art, nature, religious art and science. A system administrator of the local tourist oﬃce determined membership degrees to each category: not at all, a bit, mostly or absolutely. The system administrator can mark at most ﬁve POIs as “not to be missed”. These POIs will always be included in the initial trip that the expert system suggests to the tourist, provided that the time requirements are met.
All the data is provided and kept up–to–date by the tourist oﬃces of the diﬀerent cities. A web–based authoring tool supports this process.
6.2.2 Personal Data
The tourist ﬁrst enters a number of trip contraints that restrict the possibilities and chooses one of the ﬁve cities to visit, the date of arrival and the number of days the trip will last. Next, for each day of the trip, the tourist chooses a starting and ending location, and a start and end time. The starting and ending locations can be chosen from a predeﬁned set of locations, e.g. the railway station or a parking lot. It is also possible to enter any address, e.g. the address of a hotel. The user can request a (lunch) break of a certain duration to be scheduled in an interval during the day. E.g. a tourist prefers to have a one hour lunch break between one and three pm. Based on this information, the database is queried and only the set of POIs that can be visited within these trip constraints, based on distance and opening hours, are retrieved.
Next, the system queries the user for his or her interests and creates a user proﬁle.
This proﬁle is composed of three parts, namely types, categories and keywords. The tourist states his degree of interest in the diﬀerent categories mentioned above: not at all, a bit, mostly or absolutely. Degrees of interest should also be expressed for the POI types. Furthermore, the user may reﬁne his proﬁle by means of arbitrary keywords, e.g. football, war, Dark Ages, etc. All this does not take much more than a minute. Finally, the tourist is oﬀered the possibility to create an account in order to reuse his input for future visits to the web application.
100 THE CITY TRIP PLANNER
6.2.3 Interest Estimation
After a tourist entered the trip constraints and preferences, the system estimates the personal interest in each POI that can be visited, resulting in a score for each possible visit. This interest score is the sum of three components: the type score, the category score and the keyword search score. Every POI category membership, not at all, a bit, mostly or absolutely, is quantiﬁed as 0, 1, 2 and 3 respectively.
This is also done for the tourist’s type and category interests.
The type score of a POI is set equal to three times the interest degree of the tourist for that particular type. If the tourist is not interested in the type, the total score will be set 0, and the POI will be removed from the set of POIs. If a POI has a type score of 9, the POI is a must for the tourist and a bonus of 3 is added to the type score, in order to reinforce its possible selection. This bonus increases the attractiveness of top POIs.
For the sake of clarity, Table 6.1 presents an example of the interest score calculation of a POI, which is an abbey in this case.
In order to calculate the category score of a POI, the numerical degree of membership in each interest category is multiplied by the user’s interest for each corresponding category. Similarly, if this product is 9, an additional bonus of 3 is awarded. The category score for a POI is the sum of the three highest scoring categories, resulting in a maximum of 36.
The keyword search score is calculated using the Vector Space Model of Baeza-Yates and Ribeiro-Neto , an Information Retrieval technique widely applied by search engines. Every full text description in the POI database is preprocessed and indexed, resulting in a document vector. The user formulates interests as a number of arbitrary chosen keywords, which form a query vector. This query vector is compared to the document vector of the corresponding POI. This comparison forms the keyword search score, which is then normalised between 0 and 36. The keyword search scoring procedure is described in detail in Section 2.7.1.
The three scores are added and form the personal interest score of the tourist in the POI.
Finally, each POI that has been marked by the city’s tourist oﬃce as “not to be missed”, receives a score equal to the sum of all the POI’s interest scores plus one.
6.2.4 Tourist Trip Design Algorithm
once. A GRASP procedure performs a number of independent iterations each consisting of a constructive procedure, which is described in Section 3.5. The neighbourhood consistency algorithm (described in Section 5.3) is used when an insertion is performed. The importance of the extra constraints, however, is not taken into account when evaluating the neighbourhood. This process repeats until a ﬁxed number of iterations without improvement is reached.
In order to take the lunchbreak into account, the insertion step of Section 4.3 is extended, allowing the possibility of insertions without locations. The original insertion step tries to add, one by one, new visits to a tour. Before an extra visit can be inserted in a tour, it should be veriﬁed that all visits scheduled after the insertion place still satisfy their time window. A quick evaluation of each possible insert move is required to reduce computational eﬀorts. Checking the feasibility of all the other visits would take too much time. This can be avoided by recording W ait and M axShif t for each already included visit, as previously described in Section 4.3. A visiting duration Ti has to be taken into account. leaveid = sid + Ti is the leaving time at location i.
In order to extend this insertion step to deal with the lunch break, three cases can
be distincted when the insertion of visit j is considered between visits i and k:
1. i and j have a location, while k is a break. This case is covered by the standard formulas of Section 4.3.
When after a sequence of proposals and alteration requests, the tourist is satisﬁed with the latest proposal, the trip can be ﬁnalised. The user can print a detailed schedule of the trip, composed of all the visits with their arrival and leaving times, full text descriptions of each POI and detailed maps of the route between the POIs.
This information can also be downloaded to a mobile GPS device, allowing the tourist to track his route accurately. Moreover, the ﬁnal trip can be shared with family and friends using social networking sites.
The expert system is implemented as a web application. All data is stored in a Postgres database with spatial PostGIS extensions. User interaction is achieved via dynamic web pages using PHP and Java scripting, in an Apache web server.
The score prediction and tourist trip design algorithm was coded in Java 1.6 and exposed to the web server via a web service.
In order to demonstrate the diﬀerent planning steps of the City Trip Planner expert system, a short planning case is discussed and illustrated. A one day trip to Bruges is planned, starting from the railway station and ending in the centre of Bruges at the “Markt”. A 60 minutes lunch break is requested between noon and 2 pm (Figure 6.3).
Concerning the categories and POI types, a large interest in “monumental buildings” and “city palaces” is entered together with no interest in “statues” or “city parks” (Figure 6.4). The tourist, in this example, has a large interest in “architecture” and “local market and streets” and inputs “music” as a keyword.
The resulting trip proposal is presented in a table (Figure 6.5) and on a map (Figure 6.6). In this case, the trip contains no statues or city parks, but a number IMPLEMENTATION 103
of squares, architectural buildings and city palaces. The POI with the highest score is the “Concert Hall”. It is strongly related to the given keyword, to architecture and to monumental buildings. Although only a small interest in beguinages was given, a visit to the “Beguinage–inside” is included, since the local tourist oﬃce indicated this visit is a “not to be missed” for a tourist. Obviously, the tourist can still decide to omit this visit from the trip plan.
Figure 6.6 presents part of the map of the trip, i.
e. the end of the trip and the arrival at the “Markt”. After ﬁnalising the trip, the user can download or print a complete map with detailed POI information.
6.4 User satisfaction
6.4.1 Usage Statistics The data is collected by means of Google Analytics1 ; an on–line tool providing traﬃc statistics to a web site. Data is collected for a period of two months after the public launch of the system (July 7 till September 6, 2009).
17,510 unique visitors used the system 21,498 times, resulting in an average of 1,946 visitors per week or 282 visitors per day. Visits originated from 87 diﬀerent countries, half of which from Belgium, 20% from the Netherlands, 7% from Spain and 6% from the UK. The United States, Germany, France, Italy, Japan and Canada complete the top ten. 80% of the users visited more than one page and can thus be considered as “real” users. Only 3% leaves the site while completing one of the interest measuring forms. Two third of the users starting the trip planning wizard, also complete it. 20% of the tourists receiving a trip suggestion, decided to use the print or download functionalities. It is clear that many users just try out the web site or use it to get a ﬁrst idea about the touristic opportunities of the city. The local tourist oﬃces are enthusiastic about the number of visitors of the expert system.
An average visit took four minutes and 19 seconds and was composed of 8.07 page views. This is rather high in comparison to other web sites, which experience a lot of accidental visits. Inputting user data takes on average one minute and a half.
The user spends two minutes on average checking his personal trip proposal and reading the detailed POI information.
20,395 trips were proposed in total. In descending order, 7,313 were planned in Antwerp, 4,098 in Bruges, 3,335 in Leuven, 2,904 in Ghent and 2,745 in Mechlin.
1 http://www.google.com/analytics, last accessed November 23, 2009 106 THE CITY TRIP PLANNER 15,748 were single–day trips, 2,067 were composed out of two days, 1,213 of three days, 289 of four days, 107 of ﬁve days, 21 of six days and 143 of a whole week.
Averaged over all trips, one day of a trip consists of 26 activities on average, including start, lunch break and end. Many of these activities are only short visits (ﬁve minutes or less) to look at a statue or at the facade of a building.
6.4.2 User Feedback The website allows the users to give feedback using a questionnaire. First, the user can disagree completely with, disagree with, agree with, agree completely with or
have no opinion about the following statements:
S1 Completing the forms to construct my proﬁle took too much time;
S2 The whole process to receive a personal trip took too much time;
S3 I had to change a lot before receiving a satisfying trip;
S4 The proposed attractions meet my interests;
S5 It was always immediately clear what was expected from me;
S6 The City Trip Planner is very clear and easy to use.
Next, the user is asked for remarks, suggestions for extra functionalities, or other suggestions. During the ﬁrst two months, 43 visitors have ﬁlled in this form, of which 20 reported small technical malfunctions right after the launch. The results of the other 23 are summarised in Table 6.2.
(very) positive about the required input (S1), the response time (S2), the quality of the proposal (S3 and S4) and the user friendliness (S5 and S6). It is interesting to note that most people disagree with S1, as Davies et al.  reported that a time–comsuming input of preferences is a source of negative reactions when proposing custom–built tours.