Personalized Group Itinerary Recommendation using a

Knowledge-based Evolutionary Approach

Farzaneh Jouyandeh Pooya Moradian Zadeh
School of Computer Science School of Computer Science
Windsor, Canada Windsor, Canada
[email protected] [email protected]

ABSTRACT 1 INTRODUCTION
The problem of recommending a group itinerary is considered to A travel itinerary is a detailed planned trip schedule that consists
be NP-hard and can be defined as an optimization problem. The of the ordered list of events and locations that the traveler wants
goal is to recommend the best series of points of interest (POIs) to a to visit [14]. This plan helps travelers and tourists to get the most
group of people who are visiting a destination based on their pref- benefits from their trip by optimizing their time and costs. Although
erences and past experiences. This paper proposes an evolutionary valuable information about the destination is available on the Inter-
approach based on cultural algorithms to address this problem. Our net and travel guides, these resources can only recommend popular
objective is to maximize the group’s satisfaction by recommending points of interest (POIs), activities, events, or a generic itinerary [9].
an itinerary comprised of the optimal series of visiting POIs, consid- People’s choices, priorities, interests, and restrictions are varied,
ering the interests of all members, total travel time, and visit dura- which makes generic travel itineraries less optimized and effective.
tion while minimizing the travel costs within their assigned budget. Consequently, generating personalized itineraries based on the in-
The proposed algorithm uses historical and normative knowledge dividuals’ preferences and limitations provides more value and a
to create a belief space used later to guide the search direction higher satisfaction level.
and decision-making. The belief space is a knowledge repository Normally, itinerary planning is a time-consuming and compli-
that tracks the evolution of decisions during the search process. cated process [17]. People should search and collect information
We evaluated the performance of the proposed algorithm on a set from various data sources, process them, and select the best POIs
of real-world datasets and compared that with state-of-the-art ap- that meet their preferences and expectations [5]. Aside from POI
proaches. We also conducted non-parametric tests to analyze the selection, the second challenge is to find the best combination of
results. Compared with other algorithms, the proposed approach is these POIs and their order of visits based on travelers’ specific con-
capable of recommending efficient and satisfactory itineraries to straints, such as budget and time. This touring problem appears to
groups with diverse interests. be computationally intractable and complicated [5]. Even once a
visitor has selected an appropriate selection of POIs to visit, plan-
CCS CONCEPTS ning the proper order to visit the POIs will take a lot of time and
effort [7].
· Computing methodologies → Planning for deterministic
The itinerary recommendation problem is to autonomously rec-
actions; Heuristic function construction; · Information systems
ommend a trip plan to an individual or a group based on their
→ Expert systems.
interests and considering some constraints, including time limits,
budget, and the number of places to visit [7]. It directly links to the
KEYWORDS orienteering problem, where the underlying mathematical models
Itinerary recommendation, Cultural algorithms, Evolutionary algo- consider various constraints and satisfaction metrics when gener-
rithms, Recommendation Systems ating a path through a set of nodes [18].
This paper focuses on the personalized itinerary recommenda-
ACM Reference Format: tion problem for a group of individuals with diverse interests who
Farzaneh Jouyandeh and Pooya Moradian Zadeh. 2023. Personalized Group
wish to visit a specific destination together. Each person in a group
Itinerary Recommendation using a Knowledge-based Evolutionary Ap-
proach. In Genetic and Evolutionary Computation Conference Companion
has distinctive characteristics, interests, expectation, and travel
(GECCO ’23 Companion), July 15ś19, 2023, Lisbon, Portugal. ACM, New York, history. Additionally, the group has certain limitations to consider,
NY, USA, 9 pages. https://doi.org/10.1145/3583133.3596345 including total travel time and budget constraints. The goal is to
maximize the satisfaction of the group members based on their
preferences and limitations.
Permission to make digital or hard copies of all or part of this work for personal or Recently, this problem has attracted the attention of researchers
classroom use is granted without fee provided that copies are not made or distributed from different countries, and several approaches have been pro-
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the posed to deal with it. However, due to its natural complexity, it
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or is still an open problem. Accordingly, we approach this NP-Hard
republish, to post on servers or to redistribute to lists, requires prior specific permission problem [1] as an optimization problem with the primary objec-
and/or a fee. Request permissions from [email protected].
GECCO ’23 Companion, July 15ś19, 2023, Lisbon, Portugal tive of maximizing user satisfaction while minimizing their costs.
© 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM. To address this problem, we propose utilizing the weighted sum
ACM ISBN 979-8-4007-0120-7/23/07. . . $15.00
https://doi.org/10.1145/3583133.3596345

1684
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal Farzaneh Jouyandeh and Pooya Moradian Zadeh

method, which is one of the widely used approaches to solve these a GA that despite certain limitations, was able to generate a satis-
types of optimization problems [12]. factory travel itinerary that included a selection of highly ranked
Evolutionary Algorithms have proven to be efficient for solving tourist attractions and restaurants [19].
these optimization problems [3]. Generally, in evolutionary algo- The authors of [11] created the PersTour system, which was
rithms such as Genetic algorithms, an initial solution is evolved capable of suggesting POIs that are likely to appeal to a tourist and
through different iterations during the search process to identify a arranging them into a tour itinerary. The system offers recommen-
near-optimal solution [23]. In this paper, we propose a knowledge- dations for tours, based on either the popularity of the POIs or the
based framework based on an evolutionary model named cultural tourist’s personal interest preferences. To develop PersTour, the
algorithm (CA) [13]. CA is an extension of the Genetic Algorithm researchers utilized an adapted version of the Ant Colony Optimiza-
with a dual inheritance mechanism. In fact, as shown in Fig. 1, in tion algorithm [11]. It reinforces the selection of a single option
addition to the population space, it uses a knowledge repository over time, eventually leading to the optimal path being chosen.
named belief space. CA extracts different sources of knowledge The algorithm proposed in [22] has been designed to suggest
from the structure of the best solutions in each iteration and uses an itinerary consisting of a sequence of POIs to be visited within
them to create a new set of solutions. Using the extracted knowl- a city while maximizing the number of mandatory POIs that can
edge to guide the search direction, the algorithm can significantly be visited within the time available for travel. The authors defined
reduce the search space and find a near-optimal solution in fewer an objective function based on mandatory POIs, visit duration, and
iterations. The main advantage of using this approach is the fact total POI profit and assigned different priority levels to each of
that not only the solution is evolving during the search, but also the these metrics.
evolution of the population of solutions is trackable. Monitoring In [21], an Adaptive Genetic Algorithm (AGAM) was suggested
the evolution of belief space, a knowledge repository that stores the to address the personalized itinerary recommendation task as a
state of knowledge in each iteration, helps us observe and analyze multi-objective optimization problem. Dynamic crossover and mu-
the evolution of solutions during the optimization process. tation probabilities were incorporated to help the algorithm locate
the optimal solution and avoid getting stuck in a local best solution.
Deep learning techniques struggled to simultaneously accom-
modate numerous conflicting close and long-distance preferences,
as well as recent and prior visit influences. While some deep learn-
ing approaches, such as those based on Long Short-Term Memory
(LSTM) or Recurrent Neural Networks (RNN), were designed to ad-
dress recent visits and nearest preferences based on spatiotemporal
relationships, they still had limitations in handling these competing
factors. Therefore, learning spatiotemporal dependencies can be
complicated [2]. In [2], POI queuing time was considered for the
first time, and a Transformer-based Learning Recommendation was
proposed as a multi-task, multi-head attention transformer model.
Figure 1: The framework of the Cultural Algorithm Using two concurrent joint learning processes, it suggested the
following POIs to the target users and predicted queuing time to
gain access to the POIs. It focused solely on queuing time-aware
The remainder of this paper is organized as follows. Section 2 POI recommendations and could not provide a complete itinerary
provides a literature review, and section 3 contains the problem considering time and budget constraints.
statement and definitions. Our methodology is described in section In [7], an algorithm was designed to recommend personalized
4, followed by the experiments in section 5, and results and analysis travel itineraries that take into account users’ preferences, time
in section 6. We then provide a conclusion and our future plan. constraints, and the popularity of POIs. The Orienteering problem
was used as a framework for modeling the problem, incorporating
2 LITERATURE REVIEW factors such as time limitations and the requirement to begin and
In order to address the problem of itinerary planning effectively, end at specific POIs. Additionally, the paper introduced the concept
multiple meta-heuristic algorithms have been utilized for population- of time-based user interest, where a user’s interest in a particular
based optimization, including some evolutionary algorithms. One POI category is determined based on their duration of visit relative
such algorithm is the Genetic Algorithm (GA), which leverages nat- to the average visit duration of all users. In [10], the tour itineraries
ural selection and genetic principles to solve optimization problems. were improved by incorporating unique POIs visit duration that is
Various versions of this algorithm have been employed to tackle based on individual users’ preferences. The authors also upgraded
such challenges. For instance, in [19], the authors applied the GA the initial time-based user interest by giving more weight to recent
algorithm to the problem of travel itinerary planning with a focus POI visits and disregarding POI visits that occurred a long time ago.
on selecting appropriate restaurants. The aim of the study was to Meanwhile, in recent years, cultural algorithms have been suc-
optimize the selection of tourist attractions and restaurants in a cessfully utilized to solve some similar social network analysis prob-
travel itinerary such that the total collected utility in each location lems, such as community detection and team formation [15, 16, 24].
is maximized while staying within the specified time constraint. To Generally, similar to other evolutionary approaches it starts by
address this as an optimization problem, the researchers created creating an initial population [13]. After that, the performance of

1685
Personalized Group Itinerary Recommendation using a Knowledge-based Evolutionary Approach GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal

each individual in the population is evaluated by a fitness function. Table 1: Main constraints for the group travel itinerary
The individual in the population will then be sorted based on their Limit Description
fitness scores. In this step, different types of knowledge (e.g., nor- Visiting POIs No POI is visited more than once
mative, temporal, situational) are extracted from the structure of a The time taken for the itinerary is within the
selected group of individuals with better performance and stored Time Limit
time limit MaxT
in a knowledge repository called belief space. The belief space will Maximum number of POIs in an itinerary
The number of POIs
then play an essential role in guiding the search direction by in- should not exceed MaxPOI
fluencing the generation of the new population in the subsequent Budget
The entrance cost of the POIs should not
iterations. The process repeats until the termination conditions exceed the budget MaxB
are met. We believe that this approach can an effective solution to
address the itinerary recommendation problem.
the number of POIs in the itinerary does not exceed MaxPOI. The
3 PROBLEM STATEMENT group also has a budget (MaxB) in which the entrance cost of
recommended POIs should not be more than that, which is provided
Assume a region is represented by a complete weighted graph
in constraint 3. Finally, we ensure that the itinerary does not contain
𝐺 = (𝑉 , 𝐸, ℎ), where 𝑉 = {𝑣 1, 𝑣 2, ..., 𝑣𝑚 } is the set of m POIs, and
any duplicated POIs.
each edge 𝑒𝑖 𝑗 ∈ 𝐸 represents the route from 𝑣𝑖 to 𝑣 𝑗 . The weight
In the first step, a population of size n is created which includes
of each 𝑒𝑖 𝑗 presented by ℎ𝑖 𝑗 denotes the distance between 𝑣𝑖 and
n randomly generated solutions from the set of POIs. The popula-
𝑣 𝑗 . Each POI is assigned to a category 𝑐𝑘 (such as amusement,
tion is represented by an "n by MaxPOI" matrix, where each row
cultural, etc.), where 𝑐𝑘 ∈ 𝐶 = {𝑐 1, 𝑐 2, ..., 𝑐𝑧 }, 2 ≤ 𝑧. A set of d users
represents a generated itinerary. Therefore, let 𝑃 = {𝐼 1, 𝐼 2, ..., 𝐼𝑛 }
forms a group denoted by 𝑈 = {𝑢 1, 𝑢 2, ..., 𝑢𝑑 } who wants to visit
represent a set of n potential solutions for a problem, where each
this region. Each user 𝑢𝑖 has a category interest set represented
solution is demonstrated by an itinerary 𝐼𝑖 = [𝑣𝑢 , ..., 𝑣𝑘 ], 𝑣 𝑗 ∈ 𝑉
by 𝐻𝑢𝑖 = {(𝑐 1, 𝑣𝑎𝑙 1 ), (𝑐 2, 𝑣𝑎𝑙 2 ), ..., (𝑐𝑘 , 𝑣𝑎𝑙𝑘 )}, where 𝑣𝑎𝑙𝑡 is a value
consists of a sequence of POIs with the maximum length of MaxPOI.
between 0 and 1 that shows users’ interest level to category 𝑐𝑡 . Also,
The following example in Fig. 2 shows a generated population P
𝑣𝑎𝑙𝑡 is extracted from users’ travel history and calculated based on
with a MaxPOI of 4. Each itinerary contains up to 4 POIs and if an
their spent time on POIs of each category. Therefore, the higher
itinerary has less than MaxPOIs, the rest of the POIs are set to zero.
𝑣𝑎𝑙𝑡 means the user has a higher level of interest in its associated
category. In addition, each group has a specific time to complete the
trip denoted by 𝑀𝑎𝑥𝑇 , a 𝑀𝑎𝑥𝑃𝑂𝐼 limit on the number of visiting
POIs, and a specific budget denoted by 𝑀𝑎𝑥𝐵.
The problem is defined as identifying the most suitable itinerary
for a group of users considering their interests based on their travel
history and the group’s constraints. The main goal is to find the
optimal solution that achieves the group’s highest satisfaction level
while not exceeding its limitations. As an example, imagine there is Figure 2: An example of a population structure
a group of users consisting of three tourists with some limitations,
such as budget ($1000), time limit (8 hours), and the maximum num-
4.1 Fitness function
ber of visiting POIs (4 POIs). This group aims to travel to a region
that has twenty pre-defined POIs {𝑣 1, 𝑣 2, ..., 𝑣 20 }. The recommended After generating the initial population, the quality of all itineraries
itinerary for this group is formed as a list of the selected POIs, such is evaluated using a fitness function, and a score is assigned to each
as [𝑣 1, 𝑣 3, 𝑣 7, 𝑣 5 ]. It depicts the points to visit and their sequence. of them. The fitness score is calculated using group interests, the
total number of visiting POIs, the total popularity, visit duration,
4 PROPOSED METHOD and entrance costs.
First, each itinerary, 𝐼𝑖 , is evaluated to check whether it exceeds
Our proposed method for recommending a trip itinerary to a given
the group limitations, including time and budget. If it breaks the
group is based on a CA framework, and it takes into account
constraints, the fitness score, 𝐹 (𝐼𝑖 ), is set to -1. Otherwise, it is
four main trip constraints as shown in Table 1. Therefore, if 𝐼 =
calculated using the following formula [21]:
[𝑣𝑢 , ..., 𝑣𝑘 ] represents a recommended travel itinerary, the following
conditions should be met: 𝐹 (𝐼𝑖 ) = 𝑤 1𝐺𝑟𝑜𝑢𝑝𝐼𝑛𝑡𝑈 (𝐼𝑖 ) + 𝑤 2𝑇 𝑛(𝐼𝑖 )+
𝑘
(4)
∑︁ 𝑤 3𝑇 𝑝 (𝐼𝑖 ) + 𝑤 4𝑇 𝑣 (𝐼𝑖 ) + 𝑤 5 (1 − 𝑇𝑐 (𝐼𝑖 ))
𝑉 𝑖𝑠𝑖𝑡𝐷𝑢𝑟𝑎𝑡𝑖𝑜𝑛(𝑣𝑖 ) ≤ 𝑀𝑎𝑥𝑇 , ∀𝑣𝑖 ∈ 𝐼 (1)
𝑖=𝑢
Where 𝑤 𝑗 is the weight of each factor that can be adjusted to
|𝐼 | ≤ 𝑀𝑎𝑥𝑃𝑂𝐼 (2)
∑︁𝑘 change the impact of it on fitness score. These weights are re-
𝐶𝑜𝑠𝑡 (𝑣𝑖 ) ≤ 𝑀𝑎𝑥𝐵, ∀𝑣𝑖 ∈ 𝐼 (3) ceived as input parameters in our model and determined manually.
𝑖=𝑢 𝐺𝑟𝑜𝑢𝑝𝐼𝑛𝑡𝑈 (𝐼𝑖 ) is the total interest of the group 𝑈 to the POIs in
As formulated in constraint 1, the time that the itinerary takes to 𝐼𝑖 . 𝑇 𝑛(𝐼𝑖 ) is the total number of POIs included in 𝐼𝑖 , and 𝑇 𝑝 (𝐼𝑖 ) is
complete should be less than the group time limit (MaxT). Moreover, the total popularity of the points included in 𝐼𝑖 . 𝑇 𝑣 (𝐼𝑖 ) is the total
the group will give the maximum number of POIs they are capable duration of the visit, which is the total time that 𝐼𝑖 takes to complete.
of visiting in advance (MaxPOI). Thus, constraint 2 ensures that Finally, 𝑇𝑐 (𝐼𝑖 ) is the total cost of 𝐼𝑖 .

1686
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal Farzaneh Jouyandeh and Pooya Moradian Zadeh

𝐺𝑟𝑜𝑢𝑝𝐼𝑛𝑡𝑈 (𝐼𝑖 ) is a normalized sum of users’ interest for all cat-

egories in the generated itinerary divided by the number of users 𝐼 11 𝐼 12 ··· 𝐼 1𝑀𝑎𝑥𝑃𝑂𝐼 
1
in the group. It is calculated using the following equation. The 𝐼 𝐼 22 ··· 𝐼 2𝑀𝑎𝑥𝑃𝑂𝐼 
2
symbols used in the following equations are defined in Table 2. 𝐵𝑆 1 =  . .. .. ..  (7)
 .. . . . 
𝑦
𝑑 ∑︁  
∑︁ 𝐼 1
1 𝑥 𝐼𝑥2 ··· 𝐼𝑥𝑀𝑎𝑥𝑃𝑂𝐼 
𝐺𝑟𝑜𝑢𝑝𝐼𝑛𝑡𝑈 (𝐼𝑖 ) = (𝐼𝑛𝑡𝑢 𝑗 (𝑐𝑧 )𝛿 (𝐶𝑎𝑡 (𝑣𝑧 ) = 𝑐𝑧 )),
|𝑈 | 𝑧=1 𝑗=𝑠 (5) The sequence of events is crucial to this problem; therefore, the
∀𝑐𝑧 ∈ 𝐶 second section of the belief space, 𝐵𝑆 2 , reserves the order of POIs
in the selected itineraries. The frequency that each POI occurred
where 
1 if 𝐶𝑎𝑡 𝑣𝑧 = 𝑐𝑧 after each of the other POIs in 𝐵𝑆 1 is saved in 𝐵𝑆 2 . The historical
𝛿 (𝐶𝑎𝑡 𝑣𝑧 = 𝑐𝑧 ) = knowledge is extracted and collected in this matrix. Assume having
0 Otherwise.
a set of m POIs in a region, 𝐵𝑆 2 is formed as follows:

Table 2: Symbols definition  0 𝑓 𝑟 12 𝑓 𝑟 13 ··· 𝑓 𝑟 1𝑚 

 
 𝑓 𝑟 21 0 𝑓 𝑟 23 ··· 𝑓 𝑟 2𝑚 
Symbol Meaning 
 . .. .. .. .. 
𝐶𝑎𝑡 ( 𝑣𝑧 ) Category of POI 𝑣𝑧 𝐵𝑆 2 =  .. . . . .  (8)

𝑁 𝑜𝑝 (𝐼𝑖 ) The number of POIs included in 𝐼𝑖 𝑓 𝑟 1 2
𝑓 𝑟𝑚−1 3
𝑓 𝑟𝑚−1 ··· 𝑓 𝑟𝑚−1 
𝑚
 𝑚−1
|V| Total number of POIs  𝑓 𝑟1 𝑓 𝑟𝑚 2 𝑓 𝑟𝑚 3 ··· 0 
 𝑚
𝑇 𝑜𝑡𝑝 (𝐼𝑖 ) Total popularity of the POIs in 𝐼𝑖
MaxP The largest popularity among all POIs where 𝑓 𝑟 𝑖𝑗 is the number of times POI 𝑣𝑖 has been located right
𝑇 𝑜𝑡 𝑣 (𝐼𝑖 ) Total visit duration and travel time for 𝐼𝑖 after 𝑣 𝑗 in 𝐵𝑆 1 .
MaxV The longest visit duration among all POIs Using this belief space, the method identifies a pattern among
𝑇 𝑜𝑡𝑐 (𝐼𝑖 ) Total entrance cost for POIs in 𝐼𝑖
the best solutions of the current generation. In fact, the belief space
MaxC The largest entrance cost among all POIs
looks into the common attributes and patterns that distinguish
𝐼𝑛𝑡𝑢 𝑗 (𝑐𝑧 ) Interest of user 𝑢 𝑗 to POI category 𝑐𝑧
the best solutions of the current generation from the rest of the
itineraries in the population. The proposed structure enables us to
As shown in the following equations (6), all the other factors of take the order of POIs into account in addition to the POIs selection.
the fitness function are divided by the maximum value of its kind In the subsequent iterations, the new population will be gen-
to normalize all the values (To determine the highest value of each erated mainly using the extracted knowledge stored in the belief
factor, we utilize the available records for the specific region we space. This reduces the search space and accelerates the evolution
are targeting. For example, to normalize 𝑇 𝑛, we divide it by the process. A relatively small portion of itineraries (e.g., 20% of them)
number of POIs in the region, which is equivalent to the size of 𝑉 .). is generated using classic crossover and mutation operations. These
The number of POIs in itinerary 𝐼𝑖 is divided by the total number of operators help the algorithm to escape from the local maxima.
POIs in the region. 𝑀𝑎𝑥𝑃 is calculated as the maximum popularity In order to generate the next population, the first POI of the
that an itinerary of size 𝑀𝑎𝑥𝑃𝑂𝐼 can provide. 𝑀𝑎𝑥𝑉 and 𝑀𝑎𝑥𝐶 are new itinerary is randomly selected from the first column of 𝐵𝑆 1 . In
the longest possible duration and the highest cost of an itinerary order to select the second POI, the algorithm looks into the 𝐵𝑆 2 and
of size 𝑀𝑎𝑥𝑃𝑂𝐼 , respectively. selects the POI that has the highest frequency value. This process
continues until the MaxPOI reaches. Assuming 𝑣𝑖 is selected from
𝑁𝑜𝑝 (𝐼𝑖 ) 𝑇𝑜𝑡𝑝 (𝐼𝑖 )
𝑇 𝑛(𝐼𝑖 ) = , 𝑇 𝑝 (𝐼𝑖 ) = 𝐵𝑆 1 for the first node, for picking the next one, the algorithm looks
|𝑉 | 𝑀𝑎𝑥𝑃
at the 𝑣𝑖 -th row of 𝐵𝑆 2 . First, it chooses the columns with the highest
(6) values, which means the POI with a high frequency of occurrence
𝑇𝑜𝑡𝑣 (𝐼𝑖 ) 𝑇𝑜𝑡𝑐 (𝐼𝑖 ) after 𝑣𝑖 . If there is only one node with the highest value in that
𝑇 𝑣 (𝐼𝑖 ) = , 𝑇𝑐 (𝐼𝑖 ) =
𝑀𝑎𝑥𝑉 𝑀𝑎𝑥𝐶 row, it is selected as the next POI. In case of more than one column
After calculating the fitness score for all the itineraries, the pop- with the highest value, one of them is randomly selected. Thus,
ulation is sorted based on fitness score. Consequently, the top x% if the selected value is located in the 𝑣 𝑗 -th column, then 𝑣 𝑗 is the
of itineraries with the highest fitness values will be selected. next POI on that itinerary. Algorithm 1 shows the pseudo-code of
our approach for finding the set of POIs and their sequence as the
4.2 Belief space best itinerary to recommend. The process of extracting knowledge
The core of our proposed approach is its belief space, a knowledge from the population and creating 𝐵𝑆 2 is outlined in Algorithm 2.
repository to store different sources of knowledge extracted from Moreover, the procedure that uses the stored knowledge to generate
the selected itineraries. We form the belief space by defining two a new solution is presented in Algorithm 3.
matrices. The first one, 𝐵𝑆 1 , is a "x by MaxPOI" matrix created from
historical knowledge. It is defined as a set of selected itineraries. 5 EXPERIMENTS AND EVALUATION
Therefore, each row represents a set of POIs found in the structure In order to evaluate the performance of our model, we conducted
of the best solutions of that generation. The structure of 𝐵𝑆 1 is a series of experiments and analyses. All the experiments were
inspired from [15, 24] and presented as follows: executed on a PC with an Intel Core I7 CPU, Windows 10, and 8GB

1687
Personalized Group Itinerary Recommendation using a Knowledge-based Evolutionary Approach GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal

Algorithm 1 Knowledge-based Itinerary Recommendation Algorithm 3 Generate From BS procedure

Input: Graph G; Time limit 𝑀𝑎𝑥𝑇 ; Budget MaxB; Population size Input : Belief space matrices 𝐵𝑆 1 and 𝐵𝑆 2 ; Maximum number of
n; Iteration number 𝛿; MaxPOI; The proportion of population to visiting POIs in an itinerary 𝑀𝑎𝑥𝑃𝑂𝐼
build belief space 𝑥; Probability of generating individual from Output: Generated individual I
belief space 𝑝𝑟𝑜𝑏1; Probability of generating individual using 1: procedure GenerateFromBS(𝐵𝑆 1 , 𝐵𝑆 2 , 𝑀𝑎𝑥𝑃𝑂𝐼 )
crossover 𝑝𝑟𝑜𝑏2 2: Initialize I as an empty array with size MaxPOI
Output: Best solution 3: 𝐼 [1] ← select a POI randomly from 𝐵𝑆 1 [1]
1: 𝑃 ← Generate n individuals randomly with size 𝑀𝑎𝑥𝑃𝑂𝐼 as 4: for 𝑗 = 2 to 𝑀𝑎𝑥𝑃𝑂𝐼 do
the initial population set, considering time and budget 5: 𝑥 ← select one column randomly from highest values
constraint: 𝐶𝑜𝑠𝑡 (𝐼 𝑗 ) ≤ 𝑀𝑎𝑥𝐵 , 𝑇𝑜𝑡𝑎𝑙𝑇 𝑖𝑚𝑒 (𝐼 𝑗 ) ≤ 𝑀𝑎𝑥𝑇 where of 𝐵𝑆 2 [𝐼 [ 𝑗 − 1]]
𝐼 𝑗 ∈ 𝑃; 6: 𝑛𝑒𝑤𝑃𝑂𝐼 ← select column number of x
2: for 𝑖 = 1 to 𝛿 do 7: 𝐼 [ 𝑗] ← 𝑛𝑒𝑤𝑃𝑂𝐼
3: for 𝑗 = 1 to 𝑛 do 8: return 𝐼
4: if 𝐶𝑜𝑠𝑡 (𝑃 [ 𝑗]) ≤ 𝑀𝑎𝑥𝐵 and 𝑇𝑜𝑡𝑎𝑙𝑇 𝑖𝑚𝑒 (𝑃 [ 𝑗]) ≤ 𝑀𝑎𝑥𝑇
then
5: 𝐹𝑆 ← Calculate fitness score 𝐹 (𝑃 [ 𝑗]) best setup is 70% for using belief space to create a new individual,
6: else and in other cases, 70% for using the crossover method, and 30%
7: 𝐹𝑆 ← −1 for utilizing the mutation approach. Thus, these probabilities were
8: Sort individuals in P based on their 𝐹𝑆
9: 𝑃1 = An empty array utilized for the reported results.
10: 𝑒𝑙 ← 𝑃 [1] ⊲ Keep the elite itinerary Two real-world datasets, Flickr User-POI Visits Dataset [7][8]
11: 𝑘𝑏 ← (𝑥 * length of P)/100 ⊲ Calculate x% of population and Theme Park Attraction Visits Dataset [6], were used which
length were extracted from the Yahoo! Flickr Creative Commons 100M to
12: 𝐵𝑆 1 [1, ..., 𝑘𝑏] ← 𝑃 [1, ..., 𝑘𝑏] ⊲ Transfer best itinerary to evaluate our proposed model. The datasets consist of a set of infor-
belief space
13: 𝐵𝑆 2 ← 𝐺𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝐵𝑆 2 (𝐵𝑆 1 ) mation regarding pictures taken by various users at specific points
14: for 𝑘 = 1 to 𝑛 − 1 do of interest. This includes details such as the time that a picture has
15: if 𝑟𝑎𝑛𝑑1() ≤ 𝑝𝑟𝑜𝑏1 then been taken by a user, the user information, its location, etc. From
16: 𝑛𝑒𝑤_𝐼 ← 𝐺𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝐹𝑟𝑜𝑚𝐵𝑆 () ⊲ generates a new this data, we have extracted details regarding users’ travel expe-
solution from belief space riences, interests in different categories of POI, and the estimated
17: else if 𝑟𝑎𝑛𝑑2() ≤ 𝑝𝑟𝑜𝑏2 then
18: 𝑛𝑒𝑤_𝐼 ← generate a new solution using crossover time it takes to visit each POI.
19: else Since the proposed algorithm is able to recommend itineraries
20: 𝑛𝑒𝑤_𝐼 ← generate a new solution using mutation for a group of tourists, we created various groups with different
21: 𝑃1[𝑘] ← 𝑛𝑒𝑤_𝐼 characteristics. Groups were initialized in three sizes 3, 5, and 10.
22: 𝑃 [1] ← 𝑒𝑙 Ten separate groups have been randomly generated from the list of
23: 𝑃 [2, ..., 𝑛] ← 𝑃1
users for each group size. Moreover, the population size is a variable
24: Calculate fitness scores of P and sort individuals
25: return 𝑃 [1] of the proposed cultural algorithm, which can vary in three values
of 100, 150, and 200. After comparing the results, 200 was chosen
for the population size of CA and other evolutionary algorithms.
Algorithm 2 Generate 𝐵𝑆 2 procedure In addition, our algorithm inputs the maximum number of visiting
Input : Belief space matrix 𝐵𝑆 1 ; Total number of POIs 𝑃𝑂𝐼 𝑁𝑢𝑚; POIs in a single itinerary, time limit, and budget. The values of
Maximum number of visiting POIs in an itinerary 𝑀𝑎𝑥𝑃𝑂𝐼 these parameters are as presented in Table 3. Moreover, the weights
Output: Belief space 𝐵𝑆 2
of the factors used in the fitness function (i.e., user interest, number
1: procedure 𝐺𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝐵𝑆 2 (𝐵𝑆 1 , 𝑃𝑂𝐼 𝑁𝑢𝑚, 𝑀𝑎𝑥𝑃𝑂𝐼 )
of POIs in the individual, total popularity, total visit duration, and
2: Initialize 𝐵𝑆 2 as an 2D array [𝑃𝑂𝐼 𝑁𝑢𝑚 × 𝑃𝑂𝐼 𝑁𝑢𝑚]
3: 𝑛 ← number of rows in 𝐵𝑆 1 total entrance cost), were set manually as 𝑤 1 = 0.5, 𝑤 2 = 0.25,
4: for 𝑖 = 1 to 𝑛 do 𝑤 3 = 0.1, 𝑤 4 = 0.1, and 𝑤 5 = 0.05, respectively. In fact, in our
5: for 𝑗 = 2 to 𝑀𝑎𝑥𝑃𝑂𝐼 do experiments, the more important factors, such as user interests,
6: 𝑥 ← 𝐵𝑆 1 [𝑖] [ 𝑗 − 1] have higher weight values.
7: 𝑦 ← 𝐵𝑆 1 [𝑖] [ 𝑗]
8: 𝐵𝑆 2 [𝑥] [𝑦] ← 𝐵𝑆 2 [𝑥] [𝑦] + 1
9: return 𝐵𝑆 2 Table 3: Input parameters
Parameter Value Description
MaxPOI 6 Maximum number of POIs
of RAM. For comparing the efficiency of our algorithm with the MaxT 8 Time limit (Hours)
existing baselines, we implemented them in Python language. MaxB 1000 Maximum budget (Dollars)
Multiple experiments were conducted to find the best settings
for these parameters. As for the number of iterations that the CA All the experiments were repeated 5 times to have a more reliable
algorithm repeats, 30 and 50 are examined, and 50 is selected with outcome, and the average value of these experiments were captured.
more satisfying results compared with 30. Moreover, we investi- In order to make a comparative study, we assess the performance
gated different probabilities of generating new individuals from of our proposed algorithm in comparison with six existing state-of-
the belief space and the probability of choosing between crossover the-art baseline approaches: Genetic [20], AGAM [21], PERSTOUR
and mutation. After testing various probabilities, we found that the [7], TLR [2], PersQ [6], and GreedyFitness.

1688
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal Farzaneh Jouyandeh and Pooya Moradian Zadeh

Table 4: Comparison of algorithms in terms of the fitness score on groups with sizes 3 & 5 on the Flickr dataset
Groups with size 3 Groups with size 5
City Group CA GA PersTour AGAM TLR PersQ Greedy CA GA PersTour AGAM TLR PersQ Greedy
0 3.005E-01 3.022E-01 2.955E-01 2.655E-01 2.940E-01 2.552E-01 2.733E-01 2.84E-01 2.80E-01 2.62E-01 2.62E-01 2.86E-01 2.21E-01 2.60E-01
1 2.900E-01 2.864E-01 2.828E-01 2.618E-01 2.835E-01 2.468E-01 2.800E-01 3.023E-01 3.03E-01 2.98E-01 2.77E-01 2.96E-01 2.66E-01 2.73E-01
2 2.825E-01 2.787E-01 2.638E-01 2.520E-01 2.854E-01 2.588E-01 2.417E-01 2.80E-01 2.78E-01 2.57E-01 2.39E-01 2.80E-01 2.25E-01 2.59E-01
3 2.866E-01 2.834E-01 2.765E-01 2.539E-01 2.819E-01 2.497E-01 2.182E-01 2.96E-01 2.95E-01 2.95E-01 2.61E-01 2.94E-01 2.58E-01 2.82E-01
4 2.806E-01 2.788E-01 2.573E-01 2.472E-01 2.797E-01 2.157E-01 2.596E-01 2.70E-01 2.69E-01 2.46E-01 2.45E-01 2.69E-01 2.29E-01 2.48E-01
Budapest 5 2.828E-01 2.795E-01 2.696E-01 2.447E-01 2.578E-01 2.209E-01 2.452E-01 2.81E-01 2.80E-01 2.69E-01 2.53E-01 2.78E-01 2.32E-01 2.79E-01
6 3.032E-01 2.998E-01 2.922E-01 2.635E-01 3.044E-01 2.476E-01 2.992E-01 2.74E-01 2.73E-01 2.56E-01 2.44E-01 2.76E-01 2.17E-01 2.49E-01
7 2.884E-01 2.886E-01 2.755E-01 2.602E-01 2.859E-01 2.587E-01 2.487E-01 2.99E-01 2.99E-01 2.93E-01 2.59E-01 2.92E-01 2.57E-01 2.72E-01
8 2.988E-01 2.968E-01 2.763E-01 2.602E-01 2.975E-01 2.271E-01 2.743E-01 2.73E-01 2.72E-01 2.63E-01 2.47E-01 2.68E-01 2.32E-01 2.52E-01
9 3.023E-01 2.960E-01 2.940E-01 2.757E-01 2.928E-01 2.529E-01 2.759E-01 2.76E-01 2.75E-01 2.65E-01 2.48E-01 2.76E-01 2.32E-01 2.75E-01
0 2.125E-01 2.111E-01 1.711E-01 2.095E-01 1.933E-01 1.588E-01 1.513E-01 2.31E-01 2.02E-01 2.62E-01 2.62E-01 1.89E-01 1.60E-01 2.60E-01
1 2.005E-01 1.955E-01 1.697E-01 2.146E-01 1.889E-01 1.588E-01 1.949E-01 2.04E-01 1.98E-01 2.98E-01 2.77E-01 1.89E-01 1.59E-01 2.73E-01
2 2.365E-01 2.031E-01 1.699E-01 2.223E-01 1.895E-01 1.588E-01 1.519E-01 2.149E-01 1.97E-01 2.57E-01 2.39E-01 1.88E-01 1.59E-01 2.59E-01
3 2.153E-01 2.078E-01 1.709E-01 2.334E-01 1.927E-01 1.588E-01 1.993E-01 2.32E-01 1.95E-01 2.95E-01 2.61E-01 1.88E-01 1.59E-01 2.82E-01
4 2.098E-01 1.977E-01 1.702E-01 2.219E-01 1.903E-01 1.588E-01 2.187E-01 2.32E-01 2.02E-01 2.46E-01 2.45E-01 1.90E-01 1.59E-01 2.48E-01
Delhi 5 2.274E-01 1.980E-01 1.706E-01 2.450E-01 1.918E-01 1.588E-01 2.089E-01 2.34E-01 1.99E-01 2.69E-01 2.53E-01 1.89E-01 1.59E-01 2.79E-01
6 2.051E-01 2.054E-01 1.695E-01 2.226E-01 1.884E-01 1.588E-01 1.947E-01 2.16E-01 1.91E-01 2.56E-01 2.44E-01 1.89E-01 1.59E-01 2.49E-01
7 2.159E-01 1.989E-01 1.695E-01 2.295E-01 1.884E-01 1.588E-01 1.889E-01 2.03E-01 1.98E-01 2.93E-01 2.59E-01 1.88E-01 1.59E-01 2.72E-01
8 2.362E-01 1.955E-01 1.701E-01 2.126E-01 1.902E-01 1.588E-01 1.866E-01 2.15E-01 2.01E-01 2.63E-01 2.47E-01 1.89E-01 1.59E-01 2.52E-01
9 2.184E-01 2.021E-01 1.696E-01 2.143E-01 1.886E-01 1.588E-01 1.651E-01 2.39E-01 2.01E-01 2.65E-01 2.48E-01 1.88E-01 1.59E-01 2.75E-01
0 2.473E-01 2.400E-01 1.970E-01 2.204E-01 2.302E-01 1.647E-01 1.895E-01 2.32E-01 2.26E-01 2.03E-01 2.21E-01 2.40E-01 1.80E-01 2.06E-01
1 2.547E-01 2.428E-01 1.970E-01 2.147E-01 2.570E-01 1.658E-01 1.989E-01 2.28E-01 2.27E-01 1.97E-01 2.24E-01 2.23E-01 1.83E-01 2.24E-01
2 2.641E-01 2.679E-01 1.970E-01 2.376E-01 2.723E-01 2.080E-01 2.101E-01 2.34E-01 2.25E-01 2.03E-01 2.17E-01 2.44E-01 1.72E-01 2.06E-01
3 2.491E-01 2.495E-01 1.970E-01 2.211E-01 2.290E-01 2.032E-01 2.195E-01 2.18E-01 2.12E-01 1.97E-01 2.11E-01 2.26E-01 1.64E-01 1.87E-01
4 2.308E-01 2.240E-01 2.018E-01 2.153E-01 2.310E-01 1.646E-01 1.557E-01 2.45E-01 2.38E-01 1.98E-01 2.26E-01 2.42E-01 1.80E-01 2.04E-01
Glasgow 5 2.649E-01 2.659E-01 2.127E-01 2.289E-01 2.342E-01 1.668E-01 1.638E-01 2.23E-01 2.17E-01 1.98E-01 2.16E-01 2.30E-01 1.69E-01 2.15E-01
6 2.228E-01 2.158E-01 2.019E-01 2.075E-01 2.293E-01 1.653E-01 1.911E-01 2.38E-01 2.24E-01 2.16E-01 2.18E-01 2.31E-01 1.68E-01 1.50E-01
7 2.605E-01 2.625E-01 1.990E-01 2.145E-01 2.593E-01 1.772E-01 1.747E-01 2.21E-01 2.18E-01 1.97E-01 2.11E-01 2.29E-01 1.70E-01 2.08E-01
8 2.272E-01 2.200E-01 1.983E-01 2.129E-01 2.269E-01 1.647E-01 2.058E-01 2.38E-01 2.33E-01 1.99E-01 2.16E-01 2.46E-01 1.77E-01 2.32E-01
9 2.397E-01 2.323E-01 1.970E-01 2.166E-01 2.519E-01 1.656E-01 1.679E-01 2.33E-01 2.32E-01 2.17E-01 2.23E-01 2.27E-01 1.92E-01 1.81E-01
0 2.598E-01 2.593E-01 2.366E-01 2.320E-01 2.611E-01 1.774E-01 1.697E-01 2.36E-01 2.35E-01 2.11E-01 2.11E-01 2.33E-01 1.83E-01 1.86E-01
1 2.499E-01 2.573E-01 2.397E-01 2.309E-01 2.567E-01 1.899E-01 1.998E-01 2.71E-01 2.77E-01 2.20E-01 2.31E-01 2.67E-01 1.75E-01 1.97E-01
2 2.850E-01 2.864E-01 2.516E-01 2.357E-01 2.716E-01 1.948E-01 1.730E-01 2.36E-01 2.32E-01 2.16E-01 2.10E-01 2.37E-01 1.82E-01 2.02E-01
3 2.491E-01 2.412E-01 2.293E-01 2.099E-01 2.465E-01 1.727E-01 1.857E-01 2.63E-01 2.65E-01 2.16E-01 2.26E-01 2.62E-01 1.82E-01 2.06E-01
4 2.324E-01 2.277E-01 2.132E-01 2.074E-01 2.267E-01 1.958E-01 2.078E-01 2.30E-01 2.32E-01 2.31E-01 2.22E-01 2.34E-01 1.78E-01 1.96E-01
Vienna 5 2.219E-01 2.174E-01 2.218E-01 2.034E-01 2.238E-01 1.782E-01 1.841E-01 2.36E-01 2.37E-01 2.07E-01 2.22E-01 2.42E-01 1.83E-01 1.85E-01
6 2.564E-01 2.530E-01 2.072E-01 2.230E-01 2.551E-01 1.916E-01 2.232E-01 2.35E-01 2.29E-01 2.16E-01 2.04E-01 2.30E-01 1.84E-01 1.97E-01
7 2.886E-01 2.870E-01 2.544E-01 2.288E-01 2.717E-01 1.960E-01 1.853E-01 2.61E-01 2.63E-01 2.15E-01 2.17E-01 2.52E-01 1.81E-01 1.91E-01
8 2.481E-01 2.452E-01 2.125E-01 2.089E-01 2.431E-01 1.921E-01 1.958E-01 2.35E-01 2.30E-01 2.21E-01 2.10E-01 2.38E-01 1.77E-01 2.13E-01
9 2.603E-01 2.685E-01 2.338E-01 2.208E-01 2.594E-01 1.875E-01 1.969E-01 2.39E-01 2.33E-01 2.02E-01 2.08E-01 2.38E-01 1.83E-01 1.80E-01

Table 5: Comparison of algorithms in terms of the fitness score on groups with size 10 of the Flickr dataset
City Group CA GA PersTour AGAM TLR PersQ Greedy
0 2.695E-01 2.690E-01 2.583E-01 2.505E-01 2.680E-01 2.285E-01 2.699E-01
1 2.754E-01 2.740E-01 2.518E-01 2.500E-01 2.759E-01 2.226E-01 2.478E-01
2 2.813E-01 2.824E-01 2.587E-01 2.468E-01 2.828E-01 2.314E-01 2.538E-01
3 2.748E-01 2.743E-01 2.630E-01 2.485E-01 2.726E-01 2.281E-01 2.748E-01
4 2.767E-01 2.751E-01 2.572E-01 2.435E-01 2.812E-01 2.238E-01 2.559E-01
Budapest 5 2.757E-01 2.756E-01 2.656E-01 2.549E-01 2.696E-01 2.326E-01 2.499E-01
6 2.788E-01 2.778E-01 2.674E-01 2.524E-01 2.710E-01 2.336E-01 2.514E-01
7 2.702E-01 2.668E-01 2.544E-01 2.400E-01 2.698E-01 2.196E-01 2.670E-01
8 2.893E-01 2.859E-01 2.810E-01 2.560E-01 2.824E-01 2.496E-01 2.577E-01
9 2.790E-01 2.800E-01 2.683E-01 2.438E-01 2.778E-01 2.344E-01 2.792E-01
0 2.213E-01 1.920E-01 1.694E-01 2.088E-01 1.882E-01 1.588E-01 1.651E-01
1 2.104E-01 1.984E-01 1.695E-01 2.030E-01 1.885E-01 1.588E-01 1.938E-01
2 2.371E-01 1.987E-01 1.695E-01 2.275E-01 1.883E-01 1.588E-01 1.502E-01
3 2.307E-01 1.819E-01 1.695E-01 2.304E-01 1.883E-01 1.598E-01 1.658E-01
4 2.153E-01 1.965E-01 1.695E-01 2.308E-01 1.883E-01 1.588E-01 1.502E-01
Delhi 5 2.043E-01 1.854E-01 1.695E-01 2.033E-01 1.884E-01 1.594E-01 1.944E-01
6 2.106E-01 1.891E-01 1.695E-01 2.303E-01 1.884E-01 1.588E-01 1.938E-01
7 2.228E-01 1.998E-01 1.695E-01 2.162E-01 1.882E-01 1.588E-01 1.651E-01
8 2.056E-01 2.016E-01 1.695E-01 2.015E-01 1.884E-01 1.588E-01 1.938E-01
9 2.366E-01 1.893E-01 1.695E-01 2.260E-01 1.882E-01 1.588E-01 1.651E-01
0 2.377E-01 2.294E-01 1.975E-01 2.233E-01 2.480E-01 1.846E-01 2.220E-01
1 2.319E-01 2.247E-01 1.976E-01 2.176E-01 2.282E-01 1.721E-01 1.898E-01
2 2.342E-01 2.305E-01 2.002E-01 2.182E-01 2.498E-01 1.699E-01 1.593E-01
3 2.343E-01 2.261E-01 2.003E-01 2.147E-01 2.452E-01 1.654E-01 2.055E-01
4 2.446E-01 2.256E-01 2.073E-01 2.157E-01 2.435E-01 1.657E-01 1.768E-01
Glasgow 5 2.346E-01 2.303E-01 1.972E-01 2.240E-01 2.490E-01 1.781E-01 2.057E-01
6 2.426E-01 2.230E-01 1.971E-01 2.159E-01 2.412E-01 1.688E-01 2.053E-01
7 2.332E-01 2.255E-01 2.019E-01 2.065E-01 2.302E-01 1.794E-01 2.107E-01
8 2.331E-01 2.236E-01 1.971E-01 2.130E-01 2.321E-01 1.736E-01 2.203E-01
9 2.415E-01 2.200E-01 2.015E-01 2.132E-01 2.394E-01 1.678E-01 2.053E-01
0 2.403E-01 2.292E-01 2.179E-01 2.141E-01 2.392E-01 1.760E-01 1.675E-01
1 2.308E-01 2.219E-01 2.173E-01 2.099E-01 2.288E-01 1.774E-01 1.924E-01
2 2.363E-01 2.360E-01 2.199E-01 2.123E-01 2.413E-01 1.786E-01 2.000E-01
3 2.375E-01 2.363E-01 2.148E-01 2.090E-01 2.395E-01 1.770E-01 2.122E-01
4 2.390E-01 2.347E-01 2.131E-01 2.141E-01 2.344E-01 1.767E-01 1.997E-01
Vienna 5 2.399E-01 2.367E-01 2.188E-01 2.190E-01 2.321E-01 1.738E-01 2.157E-01
6 2.429E-01 2.369E-01 2.155E-01 2.098E-01 2.337E-01 1.893E-01 2.162E-01
7 2.294E-01 2.275E-01 2.032E-01 2.146E-01 2.316E-01 1.769E-01 1.978E-01
8 2.527E-01 2.543E-01 2.117E-01 2.195E-01 2.561E-01 1.762E-01 1.848E-01
9 2.510E-01 2.382E-01 2.087E-01 2.202E-01 2.444E-01 1.762E-01 2.169E-01

6 RESULTS AND ANALYSIS users’ preferences. Table 4 shows the comparison of CA and other
We ran all the algorithms on the two datasets listed above and approaches in four cities from the Flickr dataset. The cities are
compared the outcomes. The fitness score was utilized as a metric Budapest, Delhi, Glasgow, and Vienna and the size of the groups is
to assess how well the recommended travel itinerary matched the 3 and 5. Table 5 also presents the same experiment on groups with

1689
Personalized Group Itinerary Recommendation using a Knowledge-based Evolutionary Approach GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal

Table 6: Comparison of algorithms in terms of the fitness score on groups with size 3 & 5 of the Theme dataset
Groups with size 3 Groups with size 5
City Group CA GA PersTour AGAM TLR PersQ Greedy CA GA PersTour AGAM TLR PersQ Greedy
0 2.137E-01 2.156E-01 1.435E-01 1.481E-01 1.385E-01 1.606E-01 1.657E-01 1.95E-01 1.94E-01 1.47E-01 1.44E-01 1.42E-01 1.49E-01 1.61E-01
1 2.744E-01 2.769E-01 2.089E-01 2.116E-01 1.690E-01 1.922E-01 2.769E-01 1.92E-01 1.87E-01 1.49E-01 1.49E-01 1.44E-01 1.55E-01 1.88E-01
2 2.286E-01 2.186E-01 1.694E-01 1.588E-01 1.485E-01 1.693E-01 1.983E-01 2.18E-01 2.18E-01 1.62E-01 1.63E-01 1.54E-01 1.76E-01 2.18E-01
3 2.586E-01 2.586E-01 1.719E-01 1.805E-01 1.510E-01 1.874E-01 1.823E-01 2.12E-01 2.12E-01 1.57E-01 1.53E-01 1.50E-01 1.72E-01 2.12E-01
4 2.200E-01 2.181E-01 1.678E-01 1.667E-01 1.562E-01 1.782E-01 2.173E-01 2.24E-01 2.23E-01 1.61E-01 1.57E-01 1.52E-01 1.79E-01 2.23E-01
Disland 5 2.564E-01 2.554E-01 1.739E-01 1.770E-01 1.599E-01 2.065E-01 2.554E-01 2.38E-01 2.38E-01 1.66E-01 1.65E-01 1.57E-01 1.89E-01 2.38E-01
6 2.686E-01 2.691E-01 1.817E-01 1.748E-01 1.648E-01 2.164E-01 2.691E-01 2.04E-01 2.05E-01 1.60E-01 1.57E-01 1.48E-01 1.65E-01 2.06E-01
7 2.544E-01 2.538E-01 1.733E-01 1.685E-01 1.587E-01 2.053E-01 2.538E-01 2.11E-01 2.11E-01 1.58E-01 1.56E-01 1.50E-01 1.71E-01 2.11E-01
8 2.752E-01 2.607E-01 1.797E-01 1.731E-01 1.621E-01 2.140E-01 2.657E-01 2.90E-01 2.83E-01 1.87E-01 1.87E-01 1.74E-01 2.21E-01 2.83E-01
9 2.474E-01 2.374E-01 1.710E-01 1.750E-01 1.642E-01 2.007E-01 2.474E-01 2.93E-01 2.93E-01 1.90E-01 1.99E-01 1.75E-01 2.28E-01 2.93E-01
0 2.044E-01 2.029E-01 1.840E-01 1.588E-01 2.071E-01 1.460E-01 1.920E-01 2.04E-01 2.03E-01 1.84E-01 1.59E-01 2.07E-01 1.46E-01 1.92E-01
1 2.092E-01 2.120E-01 1.917E-01 1.775E-01 2.117E-01 1.487E-01 1.931E-01 2.09E-01 2.12E-01 1.92E-01 1.78E-01 2.12E-01 1.49E-01 1.93E-01
2 2.365E-01 2.348E-01 1.887E-01 1.826E-01 2.201E-01 1.727E-01 1.787E-01 2.37E-01 2.35E-01 1.89E-01 1.83E-01 2.20E-01 1.73E-01 1.79E-01
3 2.368E-01 2.293E-01 2.010E-01 1.841E-01 2.186E-01 1.625E-01 1.920E-01 2.37E-01 2.29E-01 2.01E-01 1.84E-01 2.19E-01 1.63E-01 1.92E-01
4 2.414E-01 2.414E-01 1.901E-01 1.831E-01 2.237E-01 1.468E-01 1.756E-01 2.41E-01 2.41E-01 1.90E-01 1.83E-01 2.24E-01 1.47E-01 1.76E-01
CaliAdv 5 2.652E-01 2.550E-01 2.118E-01 1.752E-01 2.282E-01 1.482E-01 2.058E-01 2.65E-01 2.55E-01 2.12E-01 1.75E-01 2.28E-01 1.48E-01 2.06E-01
6 2.354E-01 2.360E-01 2.067E-01 2.020E-01 2.163E-01 1.553E-01 1.922E-01 2.35E-01 2.36E-01 2.07E-01 2.02E-01 2.16E-01 1.55E-01 1.92E-01
7 2.301E-01 2.333E-01 2.017E-01 1.735E-01 2.184E-01 1.556E-01 2.201E-01 2.30E-01 2.33E-01 2.02E-01 1.74E-01 2.18E-01 1.56E-01 2.20E-01
8 2.948E-01 2.941E-01 2.141E-01 1.984E-01 2.415E-01 1.446E-01 1.704E-01 2.95E-01 2.94E-01 2.14E-01 1.98E-01 2.42E-01 1.45E-01 1.70E-01
9 3.044E-01 3.031E-01 2.181E-01 1.835E-01 2.451E-01 1.486E-01 1.709E-01 3.04E-01 3.03E-01 2.18E-01 1.84E-01 2.45E-01 1.49E-01 1.71E-01
0 2.423E-01 2.396E-01 2.355E-01 1.886E-01 2.403E-01 1.880E-01 2.316E-01 2.16E-01 2.18E-01 2.35E-01 1.99E-01 2.28E-01 1.87E-01 2.17E-01
1 2.955E-01 2.944E-01 2.789E-01 2.063E-01 2.804E-01 2.321E-01 2.445E-01 2.22E-01 2.13E-01 2.30E-01 1.94E-01 2.20E-01 1.93E-01 2.07E-01
2 2.319E-01 2.265E-01 2.409E-01 2.112E-01 2.424E-01 2.054E-01 2.220E-01 2.32E-01 2.32E-01 2.43E-01 2.14E-01 2.25E-01 2.04E-01 2.28E-01
3 2.525E-01 2.510E-01 2.484E-01 2.159E-01 2.516E-01 1.882E-01 1.931E-01 2.29E-01 2.28E-01 2.38E-01 2.06E-01 2.21E-01 2.04E-01 2.24E-01
4 2.342E-01 2.309E-01 2.471E-01 2.111E-01 2.284E-01 2.094E-01 2.303E-01 2.42E-01 2.36E-01 2.40E-01 2.13E-01 2.41E-01 2.06E-01 2.31E-01
MagicK 5 2.758E-01 2.758E-01 2.581E-01 2.209E-01 2.686E-01 2.144E-01 2.586E-01 2.51E-01 2.51E-01 2.49E-01 2.16E-01 2.51E-01 2.17E-01 2.41E-01
6 2.832E-01 2.828E-01 2.616E-01 2.156E-01 2.765E-01 2.235E-01 2.685E-01 2.24E-01 2.21E-01 2.35E-01 2.05E-01 2.18E-01 2.02E-01 2.17E-01
7 2.618E-01 2.656E-01 2.534E-01 2.159E-01 2.680E-01 2.174E-01 2.574E-01 2.26E-01 2.27E-01 2.39E-01 2.11E-01 2.42E-01 2.04E-01 2.23E-01
8 2.798E-01 2.772E-01 2.588E-01 2.175E-01 2.745E-01 2.220E-01 2.660E-01 2.97E-01 2.92E-01 2.69E-01 2.05E-01 2.79E-01 2.28E-01 2.73E-01
9 2.614E-01 2.611E-01 2.546E-01 2.048E-01 2.604E-01 2.127E-01 2.528E-01 3.07E-01 3.04E-01 2.73E-01 2.23E-01 2.84E-01 2.32E-01 2.80E-01

Table 7: Comparison of algorithms in terms of the fitness score on groups with size 10 of the Theme dataset
City Group CA GA PersTour AGAM TLR PersQ Greedy
0 1.977E-01 1.970E-01 1.520E-01 1.535E-01 1.461E-01 1.604E-01 1.977E-01
1 2.049E-01 2.030E-01 1.567E-01 1.549E-01 1.495E-01 1.638E-01 2.029E-01
2 1.846E-01 1.850E-01 1.466E-01 1.635E-01 1.415E-01 1.515E-01 1.846E-01
3 2.528E-01 2.528E-01 1.770E-01 1.806E-01 1.663E-01 1.975E-01 2.528E-01
4 2.384E-01 2.384E-01 1.698E-01 1.766E-01 1.610E-01 1.878E-01 2.384E-01
Disland 5 2.879E-01 2.874E-01 1.924E-01 1.888E-01 1.795E-01 2.211E-01 2.879E-01
6 1.924E-01 1.931E-01 1.501E-01 1.464E-01 1.442E-01 1.568E-01 1.924E-01
7 2.626E-01 2.588E-01 1.819E-01 1.515E-01 1.699E-01 2.040E-01 2.626E-01
8 2.099E-01 2.093E-01 1.571E-01 1.541E-01 1.505E-01 1.685E-01 2.099E-01
9 2.155E-01 2.150E-01 1.604E-01 1.625E-01 1.527E-01 1.723E-01 2.155E-01
0 2.227E-01 2.209E-01 1.987E-01 1.942E-01 2.160E-01 1.633E-01 2.112E-01
1 2.189E-01 2.247E-01 1.920E-01 1.904E-01 2.177E-01 1.500E-01 1.795E-01
2 2.046E-01 2.026E-01 1.892E-01 1.891E-01 2.111E-01 1.448E-01 1.758E-01
3 2.692E-01 2.689E-01 2.019E-01 2.008E-01 2.366E-01 1.526E-01 1.697E-01
4 2.571E-01 2.633E-01 1.961E-01 1.954E-01 2.306E-01 1.592E-01 1.797E-01
CaliAdv 5 3.057E-01 2.955E-01 2.159E-01 2.231E-01 2.486E-01 1.493E-01 1.693E-01
6 2.165E-01 2.203E-01 1.857E-01 1.846E-01 2.144E-01 1.596E-01 1.715E-01
7 2.830E-01 2.707E-01 2.057E-01 2.104E-01 2.401E-01 1.474E-01 1.694E-01
8 2.309E-01 2.279E-01 1.899E-01 1.901E-01 2.204E-01 1.469E-01 1.769E-01
9 2.296E-01 2.357E-01 1.889E-01 1.872E-01 2.232E-01 1.539E-01 1.755E-01
0 2.223E-01 2.130E-01 2.117E-01 2.011E-01 2.216E-01 1.989E-01 2.124E-01
1 2.259E-01 2.239E-01 2.364E-01 2.006E-01 2.162E-01 2.003E-01 2.159E-01
2 2.123E-01 2.125E-01 2.262E-01 1.948E-01 2.269E-01 1.918E-01 2.113E-01
3 2.714E-01 2.611E-01 2.569E-01 2.217E-01 2.614E-01 2.219E-01 2.495E-01
4 2.603E-01 2.506E-01 2.497E-01 2.115E-01 2.541E-01 2.163E-01 2.398E-01
MagicK 5 3.040E-01 3.016E-01 2.726E-01 2.283E-01 2.791E-01 2.376E-01 2.731E-01
6 2.199E-01 2.112E-01 2.298E-01 2.053E-01 2.309E-01 1.949E-01 2.089E-01
7 2.780E-01 2.763E-01 2.619E-01 2.060E-01 2.663E-01 2.263E-01 2.561E-01
8 2.392E-01 2.287E-01 2.368E-01 2.117E-01 2.317E-01 2.032E-01 2.206E-01
9 2.392E-01 2.392E-01 2.402E-01 2.047E-01 2.325E-01 2.057E-01 2.244E-01

size 10. As shown in tables 4 and 5, the outcomes indicated that In addition, Delhi was the only city where our proposed algorithm
our proposed algorithm (CA) outperformed the other baselines in outperformed all other baselines when suggesting itineraries for
72.5% of cases for group sizes of 3, 67.5% for group sizes of 5, and groups with sizes 3 and 5. This may be because Delhi does not
65% for group sizes of 10. have many cultural POIs, and therefore our proposed algorithm
By comparing the obtained results for each city, it was deter- performs more effectively when used in a scenario where any of the
mined that Budapest had the best quality solution. Upon further group members are not interested in the destination region. Upon
examination, it was discovered that the cultural category was the further examination of cases where alternative algorithms outper-
most popular among POIs in this city, which coincided with the formed CA, the TLR algorithm exhibited superior performance,
intersection of interests for some of the groups’ members. As a recommending the optimal solution in at least 25% of cases.
result, the algorithm can more easily offer a satisfying itinerary for Additionally, similar experiments were carried out on three re-
the created groups in Budapest, as there is a higher likelihood of gions from the Theme Park Attraction Visits dataset. The regions
finding cultural tourist attractions in this city compared to others.

1690
GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal Farzaneh Jouyandeh and Pooya Moradian Zadeh

evaluated were Disneyland, California Adventure, and Magic King- critical constraints such as travel time, group interests and bud-
dom. Table 6 presents the results of running CA and all the other get, and the number of POIs. Our method can solve this problem
algorithms on these regions for groups with sizes 3 and 5. Table 7 using an evolutionary approach by extracting different sources of
shows the results of a similar experiment for groups with size 10. knowledge from the structure of the top-performing solutions and
As shown in tables 6 and 7, the results revealed that our algorithm utilizing them to generate the final result.
outperformed the other methods in 73% of the experiments for The performance of our proposed method has been evaluated on
group sizes of 3, 80% for group sizes of 5, and 66% for group sizes two real-world datasets over multiple experiments with different
of 10. Although the statistics for each theme park differed from the group sizes and parameters. Its performance also was compared
others, the proposed algorithm consistently and reliably performed with six of the state-of-the-art algorithms in the field regarding the
well in all three regions and for all groups. After analyzing the quality of the recommended itineraries. The results revealed that
results of various algorithms, it was discovered that the GA method the proposed method could provide more satisfactory solutions for
had the highest performance rate among all group sizes, with a groups of different sizes. Moreover, non-parametric tests indicated
minimum of 26.6% cases, if we exclude our proposed algorithm. that the results of our algorithm significantly differ from those
This indicates that evolutionary approaches may be more effective of other algorithms. Our future work will involve exploring the
than other methodologies in solving this type of problem. In order evolution of knowledge and how group members influence each
to determine whether there is a significant difference between other. Additionally, we aim to improve our model to address the
the output of the compared algorithms, a non-parametric analysis cold-start problem for recommending an itinerary to a user with
was also performed on the findings acquired from each one. The no travel history.
Wilcoxon test [1] was used for this analysis. It is a non-parametric
method that aims to detect significant differences between the ACKNOWLEDGMENTS
means of two samples, indicating that the behaviors of the two The content of this paper is based on the thesis submitted by the first author
algorithms are dissimilar [1]. With a level of significance of 𝑎𝑙𝑝ℎ𝑎 = for her Master’s degree awarded by the University of Windsor [4].
0.05, the null hypothesis in this test is that CA and alternative
solutions yield similar outcomes with no statistically significant REFERENCES
differences. The associated results are shown in Table 8, which [1] Joaquín Derrac, Salvador García, Daniel Molina, and Francisco Herrera. 2011. A
shows that all algorithms in this table have p-values smaller than practical tutorial on the use of nonparametric statistical tests as a methodology
for comparing evolutionary and swarm intelligence algorithms. Swarm and
𝑎𝑙𝑝ℎ𝑎, which rejects the null hypothesis. Evolutionary Computation 1, 1 (2011), 3ś18.
[2] Sajal Halder, Kwan Hui Lim, Jeffrey Chan, and Xiuzhen Zhang. 2021. Transformer-
based multi-task learning for queuing time aware next POI recommendation.
Table 8: Wilcoxon statistical analysis (𝛼 = 0.05) In Pacific-Asia Conference on Knowledge Discovery and Data Mining. Springer,
GA PersTour AGAM TLR PersQ Greedy 510ś523.
[3] Yicun Hua, Qiqi Liu, Kuangrong Hao, and Yaochu Jin. 2021. A survey of evo-
Statistic 3.000 0.000 0.000 7.000 0.000 0.000 lutionary algorithms for multi-objective optimization problems with irregular
p-value 0.010 0.002 0.002 0.037 0.002 0.002 pareto fronts. IEEE/CAA Journal of Automatica Sinica 8, 2 (2021), 303ś318.
[4] Farzaneh Jouyandeh. 2022. Personalized Group Itinerary Recommendation Using
Cultural Algorithm. Master’s thesis. University of Windsor (Canada).
6.1 Discussion [5] Serhan Kotiloglu, Theodoros Lappas, Konstantinos Pelechrinis, and PP Repoussis.
2017. Personalized multi-period tour recommendations. Tourism Management
One of the potential challenges of using belief space to generate 62 (2017), 76ś88.
new populations in early iterations is that the extracted knowledge [6] Kwan Hui Lim, Jeffrey Chan, Shanika Karunasekera, and Christopher Leckie.
2017. Personalized itinerary recommendation with queuing time awareness.
may not be accurate enough to guide the search direction. While In Proceedings of the 40th international ACM SIGIR conference on research and
the mutation operator is beneficial to escape from local optima, development in information retrieval. 325ś334.
another effective strategy to deal with this problem is to dynami- [7] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Karunasekera.
2015. Personalized tour recommendation based on user interests and points
cally increase the belief space’s influence level dynamically during of interest visit durations. In Twenty-Fourth International Joint Conference on
the search process based on the state of the iteration. With this Artificial Intelligence.
[8] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Karunasekera.
approach in the early iterations, the influence of the belief space in 2016. Towards next generation touring: Personalized group tours. In Proceedings
the formation of the new populations is relatively low. However, of the International Conference on Automated Planning and Scheduling, Vol. 26.
it increases in subsequent iterations until it reaches its maximum 412ś420.
[9] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Karunasekera.
ratio. 2018. Personalized trip recommendation for tourists based on user interests,
The results of our proposed recommendation algorithm can be points of interest visit durations and visit recency. Knowledge and Information
used to make better decisions. However, reporting confidence level Systems 54, 2 (2018), 375ś406.
[10] Kwan Hui Lim, Jeffrey Chan, Christopher Leckie, and Shanika Karunasekera.
along with the recommended itinerary will enhance the decision- 2018. Personalized trip recommendation for tourists based on user interests,
making process. Additionally, we believe that exploring and ana- points of interest visit durations and visit recency. Knowledge and Information
Systems 54, 2 (2018), 375ś406.
lyzing the evolution of the belief space will create an opportunity [11] Kwan Hui Lim, Xiaoting Wang, Jeffrey Chan, Shanika Karunasekera, Christopher
to study the evolution of solutions throughout the search process. Leckie, Yehui Chen, Cheong Loong Tan, Fu Quan Gao, and Teh Ken Wee. 2016.
PersTour: A Personalized Tour Recommendation and Planning System.. In HT
(Extended Proceedings).
7 CONCLUSION AND FUTURE WORKS [12] R Timothy Marler and Jasbir S Arora. 2010. The weighted sum method for
In this paper, we proposed a knowledge-based framework for the multi-objective optimization: new insights. Structural and multidisciplinary
optimization 41 (2010), 853ś862.
personalized group itinerary recommendation problem. The goal [13] Robert G Reynolds. 1994. An introduction to cultural algorithms. In Proceedings of
is to maximize the group members’ satisfaction while considering the 3rd annual conference on evolutionary programming, World Scientific Publishing.

1691
Personalized Group Itinerary Recommendation using a Knowledge-based Evolutionary Approach GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal

World Scientific, 131ś139. International Conference on Industrial Engineering and Engineering Management
[14] Joy Lal Sarkar, Abhishek Majumder, Chhabi Rani Panigrahi, and Sudipta Roy. 2020. (IEEM). IEEE, 427ś431.
MULTITOUR: A multiple itinerary tourists recommendation engine. Electronic [20] Budhi S Wibowo and Monica Handayani. 2018. A genetic algorithm for gener-
Commerce Research and Applications 40 (2020), 100943. ating travel itinerary recommendation with restaurant selection. In 2018 IEEE
[15] Kalyani Selvarajah, Pooya Moradian Zadeh, Mehdi Kargar, and Ziad Kobti. 2019. International Conference on Industrial Engineering and Engineering Management
Identifying a team of experts in social networks using a cultural algorithm. (IEEM). IEEE, 427ś431.
Procedia Computer Science 151 (2019), 477ś484. [21] Phatpicha Yochum, Liang Chang, Tianlong Gu, and Manli Zhu. 2020. An adaptive
[16] Kalyani Selvarajah, Pooya Moradian Zadeh, Ziad Kobti, Yazwand Palanichamy, genetic algorithm for personalized itinerary planning. IEEE Access 8 (2020),
and Mehdi Kargar. 2021. A unified framework for effective team formation in 88147ś88157.
social networks. Expert Systems with Applications 177 (2021), 114886. [22] Phatpicha Yochum, Liang Chang, Tianlong Gu, Manli Zhu, and Hongliang Chen.
[17] Wouter Souffriau, Pieter Vansteenwegen, Greet Vanden Berghe, and Dirk 2020. A genetic algorithm for travel itinerary recommendation with mandatory
Van Oudheusden. 2013. The multiconstraint team orienteering problem with points-of-interest. In International Conference on Intelligent Information Processing.
multiple time windows. Transportation Science 47, 1 (2013), 53ś63. Springer, 133ś145.
[18] Wouter Souffriau, Pieter Vansteenwegen, Joris Vertommen, Greet Vanden Berghe, [23] Xinjie Yu and Mitsuo Gen. 2010. Introduction to evolutionary algorithms. Springer
and Dirk Van Oudheusden. 2008. A personalized tourist trip design algorithm Science & Business Media.
for mobile tourist guides. Applied Artificial Intelligence 22, 10 (2008), 964ś985. [24] Pooya Moradian Zadeh and Ziad Kobti. 2015. A multi-population cultural algo-
[19] Budhi S Wibowo and Monica Handayani. 2018. A genetic algorithm for gener- rithm for community detection in social networks. Procedia Computer Science 52
ating travel itinerary recommendation with restaurant selection. In 2018 IEEE (2015), 342ś349.

1692