2.1 The Orienteering Problem

Many tour recommender systems have as their starting point the Orienteering Problem [¹⁰, ²¹]. The idea orienteering problem came from a sports game which consisted of a number of checkpoints each having an associated score. Each player had to start at a given checkpoint, with a view to visit as many checkpoints as possible to accummulate scores.

The player who obtained the largest score in the smallest possible time was declared the winner. One constraint imposed was that each checkpoint had to be visited at most once. However, it was not mandatory for the player to start and end at the same point. In the past years, many researchers have been using the orienteering problem [¹¹, ²⁸] in their tour recommendation works.

2.2 Tour Recommendations based on the Orienteering Problem

In their paper, Choudhury et al. [⁷] proposed a tour itinerary based on the orienteering approach, in which the tourist begins the tour at some POI and finishes the at some other POI, where the goal was to recommend an itinerary comprising the most popular POIs, all within a given budget. Lim et al. [¹⁵] brought modifications to the orienteering problem by ensuring that the tourist visits one POI catergory he/she is interested in. Vansteenwegen et al. [²⁷] proposed an approach for adapting the tour schedule so that it would improve the overall balance between the defined degree of involvement from the starting and end, such as expenditure and all POIs. Lim et al. [¹⁷] have identified places to visit which require minimum queuing time. Algorithms have also been developed to recommend tours for groups of tourists which satisfy the different levels of interest of each tourist within the group [¹, ¹⁷].

2.3 Other Tourism Related Work

The wealth of information available in geo-tagged photos can be used to understand tourists’ behaviour and to find out how popular a given POI is. Ji et al. [¹²] use a graphical model to evaluate the popularity of a POI by using photos which have been uploaded to websites.

The amount of time spent by a tourist at a given POI and in which order he/she visits the POIs can be extracted from the images data [²⁰]. The authors in [⁶] have used geo-tagged photos to find out the location of clusters where popular and interesting activities are taking place.

In their paper Li et al. [¹⁴] were able determine the approximate location of photos [¹³]. A time aware measurement technique which considers a tourist’s current location to recommend the next POI to visit was proposed by Ying et al. [³¹].

The PIONEER algorithm suggested in this paper differs significantly from the current POI and tour recommendation schemes in that this algorithm uses geo-tagged images to categorise the interests of visitors dynamically depending on time spent at a POI and its popularity.

3.1 Local and Global Tourists

In this study, tourists are classified into two different categories - local tourists and global tourists. A tourist who has already visited a certain city is referred to as a local tourist with respect to that city. If this tourist now travels to a different city which he has not previously visited, the local tourist of that city becomes a global tourist with respect to the travelling tourist.

To make the concepts clearer, consider two tourists T1 and T2 where T1 has visited Mauritius but not Bangalore and T2 who has visited Bangalore but not Mauritius. Then T1 is a local tourist for Mauritius and T2 is a local tourist for Bangalore. Suppose now that tourist T1 wants to visit Bangalore. Then T2 becomes the global tourist for T1. Similarly, if T2 flies to Mauritius then T1 becomes the global tourist for T2.

3.2 Travelling History of a Tourist

Let U be a set of tourists. Suppose there is some tourist u∈U who has so far visited k POIs. Then, the travelling history of u is given by a sequence of triples ℍu=((p1,tp1a,tp1d),…,(pk,tpka,tpkd)).

In the triple (pι,tpιa,tpιd), pι is the POI visited by the tourist, tpιa is the time of arrival at POI pι and tpιd is the time of departure from pi. The difference between tpιa and tpιd gives the amount of time spent at pi. To make the notation simpler, we will write ℍu=(p1,…,pk) instead of ℍu=((p1,tp1a,tp1d),…,(pk,tpka,tpkd)).

3.3 Travelling Sequences of a Tourist

Given some tourist, his/her travel history is broken down into several distinct travel sequences if the time difference between two consecutive POI visits is tseq hours or more. In our work we use tseq=8 as proposed by Lim in [¹⁶]. So, the travel history ℍu1 can be written as ℍu1,ℍu2,…,ℍuk, where k is the number of travel sequences.

3.4 Average Time Spent at

UNIRa=POI

For every tourist, the history of his/her past travels is known. Given this information, the equation 1 can be used to calculate the mean time spent by all tourists who have visited a specific POI p [³, ⁵]. This value is denoted by A(p):

Algorithm 1 PIONEER Algorithm 

where u={1,2,…,q}, ȷ={1,2,…,r}, Vu is the frequency of tourist’s u visit to POI p and δ(pι=p)=1 if pι=p and 0 otherwise.

3.5 Tourist Interest for POI Category

Recall that the symbol ℂ has been used to denote the set of categories of POIs and cp∈ℂ to denote the category of a POI p. Then, the interest a particular tourist u has for a particular category c of POI can be calculated using equation 2:

where δ(cpȷ)=1 if cpȷ=c and 0 otherwise. The tourist interest for the POI category c is obtained from equation 1 by calculating the time spent by a tourist u at POI category c relative to the total time spent by all the tourists. It makes sense that a tourist will stay for a longer period at a POI category in which he/she is most interested in.

3.6 Local and Global Tourist Similarity

The degree of similarity between local and global tourists can be determined based on their interests for a given destination. For two distinct tourists ux and uy, we can compute their similarity using the cosine similarity measure as shown in equation 3:

3.7 Tourist Interest for a POI

The interest a particular tourist u has for a particular POI p can be determined using the equation 4:

where A(p) (see equation 2) is the average time spent by all tourists at POI p and δ(pi=p)=1 if pi=p and 0 otherwise.

3.8 Popularity of a POI

Every POI p has a certain popularity associated with it which is denoted by Pop(p). The popularity of a POI is taken to be the number of times the POI has been visited by all tourists. More formally, Pop(p) is defined by:

where U is the set of all tourists and Φu,p is the number of times tourist u has visited POI p.

3.9 Travelling Cost

There is a cost involved while travelling from one POI to another. In previous studies the cost of travel from one POI pi to another POI pj was a measure of the time taken by the tourist to complete the trip from pi to pj.

The total cost was considered to be the total time taken for an entire tour. The problem with using time as a measure of travel cost is that travel time depends on the means of transport used.

For example if two POIs are very far apart and the tourist decides to take a flight, then the travel time will be much lower. In this paper, the total distance travelled by the tourist has been used as the travel cost and the aim is to minimise the distance travelled in an itinerary. The total cost of an itinerary I=(p1,p2,…,pN) with N POIs is given by:

where Dist(pι,pj) is the distance between POIs pι and pj which can be calculated using the Haversine formula [²⁵].

6.1 Dataset

This paper uses the YFCC100M (Yahoo! Flickr Creative Commons 100M) dataset [²⁶]. It is a huge dataset comprising 100 million photos and videos obtained from Flickr.

From the metadata about the dataset information such as the date and time and the latitude and longitude values when the photos were taken and the ids of users who took the photos can be extracted.

6.2 Baseline Algorithms

6.3 Real-Life Evaluation

Only those tourists who have completed at least two travel sequences and visited at least two categories of POIs are used to evaluate the proposed algorithm.

The method is applied to both local and global datasets [²³], as well as visitors who are comparable. We compare similar visitors in this study by looking at the top 10 associated visitors from global data sets.

For our experiments, categories of real travelling series are chosen based on the history of associated visitors in a given area. The standard evaluation metrics, that is, Precision, Recall and F1-Score have been used to test our algorithm.

6.4 Comparison of Precision, Recall and F1-Score

The proposed PIONEER algorithm performs better when compared to other baseline algorithms such as GREEPOP, TRIC, GREENEAR and RAND. Tables 1, 2 and 3 show how the PIONEER algorithm compares with other baseline approaches in terms of Precision, Recall and F1−Score values. The results show that PIONEER fares better than the baseline approaches as far as Precision, Recall and F1-Score metrics are concerned.

Recall measurements depending upon |Crec| and |Crec∩Creal| as per in Eqn. 17. Here the values of |Crec∩Creal| is better compared to the various baseline approaches which can be computed utilizing the PIONEER algorithm. Typically, the suggested PIONEER algorithm is based on local as well as global datasets, and ultimately suggests many POIs, resulting in better Recall scores for various baseline approaches.

For the PIONEER algorithm, the Precision scores are more because they are dependent on |Crec| and |Crec∩Creal| as per in Eqn. 18. We found that Crec scores vary for various baseline approaches during the analysis. The values of |Crec∩Creal| are higher for the suggested PIONEER algorithm.

The F1-Score increase in the suggested PIONEER algorithm based on Precision and Recall compared to other baseline approaches.