3.1 Characteristics of Local and Global Users

The Local User (LU) is a user who lives in some city A, where he/she has visited some number of POIs. Suppose the LU from city A wants to visit a new city B. Then the LU of city B is referred to as the Global User (GU) with respect to the LU of city A. [⁷] contains a detailed discussion about local and global users.

3.2 Average POI Visit Duration

Considering the travel behavior of some user u, the average visit duration for a POI can be calculated using Eqn.1:

where, j={1,2,…,ℓ}, u={1,2,…,k} and V provides the number of trips a tourist makes to a certain POI. δ(ℙj=ℙ)={​0  otherwise1  if (ℙj=ℙ). D(ℙ) is the average visit duration for a specific POI. The mean visiting time for a certain POI ℙ is referred to as ℙ [², ³].

3.3 Interest in LUs and GUs Depending on Time

Suppose Cℙ is the category of some POI ℙ. The interest of a particular tourist for some category ℂ is given by Eqn. 2:

where, δ(Cℙj)={​0,  otherwise1  if (Cℙj=ℂ). Eqn. 2 will later be used to measure tourists’ interest for some POI category ℂ with respect to the visiting times of all tourists for that POI category.

It is obvious that a visitor spends more time at a certain POI if he/she is highly interested in that POI.

3.4 History of Travel

Let U be a set of tourist. For some tourist u∈U, we define a sequence of itineraries Su=((i1,ti1ari,ti1dept),…,(in,tinari,tindept)) where n is the number of itineraries, in a triplet (iϑ,tiϑari,tiϑdept), iϑ is an itinerary, tiϑari is the time of entry and tiϑdept is the time of departure. The difference between the two time values gives the duration of itinerary iϑ. For simplicity Su=((i1,ti1ari,ti1dept),…,(in,tinari,tindept)) can be written as Su=(i1,…,in).

3.5 Itinerary Interest

POIs ℙ=(ℙ1,ℙ2,ℙ3,…,ℙk) can be used to form the itinerary of the tourist. An interest value is associated with each itinerary iϑ∈Su. This value can be obtained from Eqn. 3:

3.6 Itinerary Popularity

Each itinerary also has a popularity associated with it, which can be obtained using equations 4 and 5. We first obtain the popularity of a user for a POI category c, denoted by C(popl):

where poplℙj is the user’s visit frequency to POI ℙj. The visit frequency of all the users at POI ℙj is given by Φ(ℙj). The popularity of an itinerary is then given by the following equation:

where ς = {1, 2, 3, ..., ω} is the total number of categories present in i_ϑ.

3.7 Travel Costs

Travel costs are calculated by the physical distance that has been traveled along the journey. Many earlier works take the entire travel into account. But time is dependent on means of travel like taxis, trains, airlines, walks, etc.

The distance is a significant factor if the visitors want to travel multiple POIs using a broad transportation system. We reduce journey times by using quick way of transportation.

If two POIs are a long way from each other, a fast transportation method is required and the costs of transport gets increased. We consequently seek to maintain a minimal level of the entire physical distance of the trip. Travel expenses are determined by Eqn. 6:

where, (ϑ+1)<n. The double summation in Eqn. 6 gives the total distance between the POI attractions of all the itineraries present in the travel package. The overall physical radius among all POI attractions is calculated based on internal distance iϑ.

The second summation in Eqn. 6 is the external distance between iϑ and iϑ+1 and it is calculated using the physical distance between the last POI of itinerary iϑ and the first POI of itinerary iϑ+1.

3.8 LU and GU’s 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 7:

3.9 Users to Tour Group Allocation

Because the tourist group is divided into m tours, let G={G1,…,Gm} is the tourists group’s category, and Gk={g1,…,gq} indicates a kth group includes q tourists. Our goal is to establish the following groups:

The cosine similarity metric indicates how similar two users’ interests are. Gk, and for all tour groups, in Eqn. 8.

This clustering issue has been demonstrated to have optimum solutions that are NP-hard. As a result, we employ the following approach to provide approximations to the solution to this issue.

3.10 Density-based spatial clustering of applications with noise (DBSCAN)

Based on a collection of points and the computation of euclidean distance and total number of points, DBSCAN grouped points that were similar to one another.

Outliers are frequently categorised as the points in low-density zones. This is what Euclidean distance deals with:

3.11 Problem definition

In this part, we will address the recommendation for multiple itineraries by considering POIs for a specific tourist. The major objective is to enhance visitor and POI’s popularity and to decrease traveling expenses.

The optimization issue is known as the gTravelREC problem classified as the [⁹] version of the Orienteering Problem. This portion deals with the problem of different POIs for one person.

Our main objective is to maximize the interest and popularity of visitors and to reduce expenditures.

A type of orienteering issue [⁸] could be used to resolve this issue:

Eqns. 11-19 are the different formulations of the itinerary. χi,J represents the total of A in the list of ith itineraries of sizes J for all itineraries accessible.

This work aims mainly to propose several itineraries i1,i2,…,in for maximizing visitor interest and for reducing travel costs. So, the goal can be written as:

Algorithm 1 Algorithm for NSGA-II based gTravel approach. 

4.1 Dataset

In this analysis, we utilized the data provided in [⁸].

The dataset contains images and videos by Yahoo! Flickr Creative Commons 100M (YFCC100M) [¹³]. Furthermore, the YFCC100M data set provided in Table 1 was used and geo-tagged images from different areas of the globe have been obtained.

The data set comprises the photo’s meta-data. It includes visiting dates and times.

The dataset also contains data from the Geo-coordinate to identify the length among POIs. The data sets utilized in this research could be accessed from this url^fn for free.

4.2 Baseline Algorithms

As per [⁸], we have taken into account in the experimentation that all benchmark approaches begin at one POI and then choose the following POIs till the target is met.

We utilize the series of POIs to suggest different itineraries. After creating a single itinerary, we used the remaining POIs.

We assumed four hours of travel time in a given route, which might easily aid in the creation of a variety of itineraries.

4.3 Real-life Evaluation

Only tourists who have visited at least two sequences or more would be evaluated. The system is used locally and globally, and the related users were defined.

In this analysis, we can assess the corresponding visitors by selecting the top ten linked visitors of the GU’s list. From the local data set the various attributes of the relevant visitors are collected.

We pick the following matrices to compare our solution with different benchmarks. For our experiments, the series of real-life sequences are chosen depending on the prior histories of the visitors in an area a tourist desires to explore.

4.4 Comparison of Precision, Recall and F1-Score

With respect to the benchmark approaches like GPop TOURINT and GNear, the efficiency of the gTravel algorithm is maximum. Tables 2, 3 and 4 present the values of Precision, Recall and F1-Score of the gTravel algorithm and other benchmark approaches.

The results show that in comparison with benchmark approaches, the proposed gTravel algorithm performs better. The Recall scores of the gTravel algorithm is 6.86%−26.73% higher compared with other benchmark approaches (see 3). The Recall value depends on values ‖Cv‖ and ‖Crec∩Creal‖.

In case of gTravel algorithm, the value of ‖Crec∩Creal‖ is higher than the various benchmark approaches. The gTravel algorithm typically runs on two datasets, local and global, and suggests various itineraries. This results in increased Recall values. The popularity of the tour and the interest of visitors is compared with different benchmark approaches such as GPop, TOURINT and GNear.

The Precision value of the gTravel algorithm is higher compared to other benchmark approaches from 4.71%−20.49%. The precision values depend on ‖Crec‖ and ‖Crec∩Creal‖. We found through experiments that Crec values vary for various benchmark approaches.

The ‖Crec∩Creal‖ values are higher for the gTravel algorithm. The F1-Score value of gTravel algorithm is 4.07%−22.13%, higher compared to other benchmark approaches, as the F1-Score values depend on Precision and Recall.

The performance of MULTITOUR is higher than the various benchmark approaches because it considers the popularity of itineraries, the interest of tourists and traveling expenses.

However, the other benchmarks do not support many itineraries and evaluate the popularity of attractions or interest of visitors in the location.