HybridCRS-TMS: Integrating Collaborative Recommender System and

TOPSIS for Optimal Transport Mode Selection

Mouna Rekik

1,2

, Rima Grati

, Ichrak Benmohamed

and Khouloud Boukadi

2 a

Higher Institute of Transport and Logistics of Sousse, Tunisia

University of Sfax, Multimedia, Information systems and Advanced Computing Laboratory, Tunisia

Zayed University, College of Technological Innovation, Abu Dhabi, U.A.E.

Keywords:

Collaborative Filtering, K-NN, Transport Mode Selection, Hybrid Decision, TOPSIS.

Abstract:

The pervasive inﬂuence of smartphones and mobile internet has revolutionized journey planning, particularly

transportation. With navigation applications delivering real-time information, recommender systems have

emerged as crucial tools for enhancing the travel experience. This paper introduces HybridCRS-TMS, a unique

Hybrid Collaborative Recommender System for Transport Mode Selection, leveraging a dataset of 260 pas-

sengers. Through advanced data mining techniques, speciﬁcally k-Nearest Neighbors (k-NN) for collaborative

recommendations and TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) analysis

for objective evaluation, the system provides personalized transportation mode recommendations. The model

not only demonstrates exceptional performance but also showcases the synergy between collaborative and

objective decision-making approaches, contributing to efﬁcient, personalized, and well-informed travel solu-

tions. This study underscores the system’s versatility, illustrating its ability to optimize travel choices through

a hybrid recommendation framework that integrates both collaborative and objective criteria.

1 INTRODUCTION

The advent of smartphones and mobile internet has

signiﬁcantly transformed modern living, particularly

in journey planning. These technological advance-

ments extend beyond traditional domains like e-

commerce and healthcare, reaching the transporta-

tion sector. Navigation applications have liberated

travelers from the hassle of paper maps and transit

timetables, introducing a dynamic aspect to decision-

making. Empowered by real-time information, pas-

sengers can explore transportation options based

on their starting point and destination, streamlining

route-searching and enabling informed decisions. In-

tegrating these technologies not only saves time but

also enhances the overall travel experience by provid-

ing tailored transportation choices aligned with pas-

sengers’ preferences and constraints (Liu et al., 2021).

Recommender systems, crucial in this transfor-

mation, contribute by providing personalized sugges-

tions and enhancing the overall travel experience.

These systems offer various transportation options,

ensuring passengers make informed choices based on

https://orcid.org/0000-0002-6744-711X

their preferences and needs.

In the transportation ﬁeld, these systems assist in-

dividuals, including students, employees, and work-

ers, in selecting the most suitable mode of transporta-

tion, such as a taxi, shared taxi, bus, or car. They

signiﬁcantly enhance the overall transportation expe-

rience by providing personalized recommendations

based on individual preferences and requirements.

In Tunisia, where citizens face diverse transportation

options and preferences, recommender systems help

streamline decision-making and mitigate challenges

associated with navigating various transport modes.

These systems tailor suggestions to users’ needs, op-

timizing travel choices and contributing to more efﬁ-

cient and personalized transportation solutions.

Passengers in Tunisia can access various trans-

portation options, including shared taxis, buses, indi-

vidual taxis, and personal cars if public transportation

falls short. Shared taxis compete with buses, offer-

ing high speeds and frequencies despite limited ca-

pacity and sometimes chaotic organization. Individ-

ual taxis provide comfortable and fast service, though

at a higher cost. These transportation options play

a crucial role in passengers’ lives, allowing them to

choose the mode that best suits their needs in terms of

Rekik, M., Grati, R., Benmohamed, I. and Boukadi, K.

HybridCRS-TMS: Integrating Collaborative Recommender System and TOPSIS for Optimal Transport Mode Selection.

DOI: 10.5220/0012758300003753

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Conference on Software Technologies (ICSOFT 2024), pages 383-394

ISBN: 978-989-758-706-1; ISSN: 2184-2833

383

convenience, cost, and speed.

Several existing research works focus solely on

either subjective selection approaches or objective

methods without exploring the potential beneﬁts of a

hybrid selection approach. This limited scope raises

questions about the comprehensiveness of their rec-

ommendations and their ability to leverage the com-

bined strengths of both subjective and objective crite-

ria for a more nuanced and effective decision-making

process.

This paper introduces a novel two-phased recom-

mender system utilizing a collected dataset from 260

travelers. The unique two-step decision-making pro-

cess employs a collaborative ﬁltering recommender

system in the ﬁrst phase, delivering personalized rec-

ommendations based on similar users’ preferences.

What sets this work apart is the integration of a sec-

ond phase introducing the Technique for Order of

Preference by Similarity to Ideal Solution (TOPSIS)

method, adding a layer of objectivity and robustness

to transportation mode selection. Combining subjec-

tive and objective criteria, the hybrid approach en-

sures a comprehensive and effective decision-making

mechanism for transportation mode recommenda-

tions.

The paper is structured as follows: the ﬁrst section

provides an overview of the related work. Section 3

details our proposed recommender system. Finally, in

section 4, we present the conclusion of the paper.

2 RELATED WORK

Recommendation systems in the transport sector are

of exponential interest due to their ability to assist

travelers in choosing the most convenient mode of

transportation.

The study in (Sun and Wandelt, 2021) utilizes ma-

chine learning on a travel recommendations dataset

and actual mode choices in Beijing, China. The data

is sourced from Baidu’s prototype route recommen-

dation system. Users received a summarized list of

recommended transportation modes for a speciﬁc ori-

gin/destination/time request and selected their pre-

ferred mode, leading to a detailed route.

The recommender system proposed by Wu et al.

(Wu et al., 2022) introduces an incremental scanning

method incorporating multiple time windows to ex-

tract multi-scale features from user behaviors. Addi-

tionally, a hierarchical behavior structure is devised

to alleviate the computational burden associated with

large data sets. The proposed framework aims to en-

hance social beneﬁts by dynamically adjusting can-

didate modes based on real-time trafﬁc states. This

adaptation can promote public transport use, allevi-

ate trafﬁc congestion, and reduce environmental pol-

lution.

In the same context, (Arnaoutaki et al., 2021) in-

troduces a recommender system tailored for selecting

MaaS (Mobility as a Service) plans, aiding travelers

in choosing bundles of mobility services that align

with their everyday transportation needs. The rec-

ommender system ﬁlters out unsuitable plans, subse-

quently ranking the remaining options based on their

similarity to users’ characteristics, habits, and prefer-

ences.

The model proposed by (Lai et al., 2023), entitled

Balance Multi Travel Mode Deep Learning Prediction

(BMTM-DLP), applies the concept of recommender

systems to individual travel mode prediction. The

model is leveraged to extract individual travel prefer-

ences, enhancing the accuracy of travel mode predic-

tions. Additionally, introducing a focal loss function

module within the model mitigates the impact of un-

balanced categories, contributing to more robust and

balanced predictions. (Arnaoutaki et al., 2019) intro-

duces a knowledge-based recommender system that

utilizes constraint programming mechanisms. It of-

fers functionalities to capture user preferences, elim-

inate MaaS (Mobility as a Service) plans that do not

align with those preferences, and assess the similar-

ity of the remaining plans to the user’s proﬁle. The

result is a ﬁltered and ranked list of MaaS plans, en-

abling users to choose the one that best aligns with

their preferences.

In (Rodriguez-Valencia et al., 2022), research on

user satisfaction and ridership factors in public trans-

portation (PT) has been extensive. A signiﬁcant

contribution is found in a study conducted in Bo-

got

a, Colombia, utilizing Structural Equation Mod-

eling (SEM) and Multiple Indicators Multiple Causes

(MIMIC) models. The study focuses on three PT bus

subsystems, including Bus Rapid Transit, a formal-

ized bus subsystem, and a semi-formalized counter-

part. It identiﬁes latent variables such as ”condition,”

”service,” and ”safety/security” within each subsys-

tem, highlighting the varying strengths and signiﬁ-

cance of direct and indirect effects. This research pro-

vides nuanced insights into the relationships among

infrastructure, vehicles, operational attributes, and

regulatory processes, offering valuable perspectives

for decision-makers aiming to improve PT services

and aligning with the broader discourse on enhancing

user experiences in transportation systems.

In the context of intelligent transportation, indi-

viduals typically decide on their preferred transport

modes based on personal inclinations and journey

characteristics. As the transportation landscape un-

ICSOFT 2024 - 19th International Conference on Software Technologies

384

dergoes a transformative shift with the introduction

of autonomous vehicles (AVs), it becomes crucial to

understand the potential impacts on traditional mode-

choice models. In this context, (Hamadneh and Es-

zterg

ar-Kiss, 2023) explores three transport modes:

conventional cars, privately owned autonomous vehi-

cles (PAVs), and shared autonomous vehicles (SAVs).

This study employs a discrete choice modeling ap-

proach to formulate a transportation mode choice

model. A stated preference (SP) methodology is uti-

lized, collecting 306 responses in Hungary. Individu-

als exhibit variations in their willingness to use a spe-

ciﬁc transport mode based on factors such as income,

family size, and current transportation habits.

In reviewing the related work, it is evident that

existing research in transportation recommender sys-

tems predominantly falls into two distinct categories,

each emphasizing speciﬁc aspects of the decision-

making process. On one hand, a substantial body

of work concentrates on understanding and incor-

porating users’ preferences into the recommenda-

tion process. On the other hand, another signiﬁcant

strand of research focuses on the objective selection

of transport modes, emphasizing efﬁciency and prac-

tical decision-making.

While these two approaches have individually

demonstrated their effectiveness in addressing spe-

ciﬁc facets of the transportation recommendation

challenge, integrating user-centric and objective-

oriented elements remains a relatively underexplored

area. Combining insights from user preferences with

the rigor of objective decision-making methods could

yield a more versatile and adaptable recommender

system.

3 PROPOSED HYBRID

RECOMMENDER SYSTEM

Our HybridCRS-TMS recommender system intro-

duces a collaborative recommendation approach that

incorporates all essential components of such a sys-

tem. Moreover, we integrated a Multiple Criteria De-

cision Making (MCDM) method to enhance decision-

making by combining subjective and objective ele-

ments. This integration ensures a more robust selec-

tion of transportation modes, contributing to a com-

prehensive and effective decision-making process.

The collaborative ﬁltering method has gained

widespread popularity and demonstrated signiﬁcant

success in terms of accuracy. The underlying prin-

ciple of collaborative ﬁltering methods involves ana-

lyzing users’ historical ordinal feedback information

to make predictions for recommendations. In simpler

terms, the system suggests items to a speciﬁc user

based on similar users’ preferences, independent of

the features of the items themselves (Alhijawi and Ki-

lani, 2020).

The HybridCRS-TMS, illustrated in Figure 1,

leverages collaborative ﬁltering techniques to ana-

lyze user preferences, behaviors, and historical data,

providing personalized recommendations for various

transportation modes, including taxis, shared taxis,

buses, and private cars. The system incorporates user

proﬁles, historical usage patterns, and a collaborative

ﬁltering algorithm to enhance the accuracy and rele-

vance of the recommendation. Through a comprehen-

sive approach, it considers the diverse factors inﬂu-

encing transportation choices. The output of this sys-

tem is used as an input of the second decision phase

discussed in Section 3.4

3.1 Pre-Treatment of Collected Data

3.1.1 Data Collection

A meticulously crafted questionnaire was developed

to gather insights into passengers’ preferences and

proﬁles. Ensuring the acquisition of highly accurate

data was paramount for precise results. The question-

naire comprises 14 questions divided into two parts,

each designed to extract distinct yet valuable infor-

mation.

The ﬁrst part is devoted to passengers’ proﬁles,

encompassing demographic details such as age, gen-

der, socio-professional category/function, reason for

travel, and the weekly allocated budget for trans-

portation. The second part focuses on the behavioral

aspects of transport users, including the reason for

travel, departure and arrival stops, days of travel, and

the preferred mode of transportation.

To guarantee a thorough understanding of the

questions by participants, a pre-test was conducted

among a sample of one hundred individuals. This pre-

liminary step aimed to assess the clarity and relevance

of the questions and identify any potential sources of

confusion. Feedback obtained during this pre-test was

carefully analyzed and contributed to reﬁning the ﬁ-

nal questionnaire. This proactive approach optimized

the quality of responses, ensuring that the questions

were understandable and pertinent to the diverse par-

ticipants.

During the three-week study conducted at the end

of February and the beginning of March 2023, 260

individuals aged 10 and above participated, each ded-

icating approximately ten minutes to complete the

questionnaire.

It’s vital to note that the questionnaire was tailored

HybridCRS-TMS: Integrating Collaborative Recommender System and TOPSIS for Optimal Transport Mode Selection

385

Figure 1: The proposed recommender system framework.

to capture citizens’ unique characteristics and prefer-

ences in the Sousse region of Tunisia.

3.1.2 Data Preprocessing

After the data collection phase, we proceed to a

pivot phase consisting of data preparation and pre-

treatment (Mariscal et al., 2010).

Table 1 explains in detail the dataset attributes.

This table provides information on each attribute. The

”Original values” column showcases the data as ini-

tially recorded or collected, while the ”Normalized

values” column represents the same data transformed

into a standardized format. This normalization pro-

cess ensures consistency, making it easier to analyze

and compare the attributes across the dataset.

The pre-processing phase for our HybridCRS-

TMS dataset comprises various essential steps. Ini-

tially, the dataset is loaded from an Excel ﬁle resulting

from the collection phase. Missing values are handled

through imputation using mean and mode for numer-

ical and categorical features. Categorical attributes

such as ’Reason for travel’,’Days of travel’, ’Func-

tion’,’Departure stop’ and ’Arrival stop’ are one-hot

encoded to convert them into a machine-learning-

friendly format.

One Hot Encoding is the predominant coding

scheme widely employed in data representation. This

method involves comparing every level of a cate-

gorical variable against a designated reference level.

Through One Hot Encoding, a single variable with

n observations and d distinct values is transformed

into d binary variables, each having n observations

(Kedar Potdar, 2017). The ’Transport modes’ col-

umn is transformed from a comma-separated string to

a list, and MultiLabelBinarizer is utilized to manage

this list-formatted data. It is straightforward to men-

tion that each passenger may use many transporta-

tion modes. Additionally, Min-Max scaling is applied

to normalize the ’Age’ and ’Budget’ attributes. The

Min-Max normalization transforms a variable x into a

new normalized variable x’ according to the equation

′

x − min(x)

max(x) − min(x)

(1)

where

• x’ is the new normalized variable,

• x is the original variable,

• min(x) is the minimum value of x

• max(x) is the maximum value of x

In this pre-processing phase for our collected dataset,

instances associated with less frequently chosen

transportation modes are removed from the dataset.

This step is taken to streamline the dataset and focus

ICSOFT 2024 - 19th International Conference on Software Technologies

386

Table 1: Used dataset Attributes.

Attribute Original values Normalized values

Age [10..65] Scaled values [0 and 1]

Budget [3..35] Scaled values [0 and 1]

Gender [Male,Female] [ 1 : female, 0 : male]

Function Student, Worker, Trader, Retired... One-hot encoded columns

Departure and arrival stops 5 different stops One-hot encoded columns

Reason for travel Study, Health, Purchase, Leisure, Other One-hot encoded columns

Days of travel [’Monday’, ’Tuesday’,...] One-hot encoded columns

Transportation mode Bus, Taxi, shared Taxi, Car MultiLabelBinarizer

on more prevalent and representative transportation

choices. Removing instances with infrequent trans-

port mode selections aims to enhance the dataset’s

overall quality and relevance for subsequent analyses.

3.2 Collaborative Filtering System

In this ﬁrst decision phase, we employ a collaborative-

based recommender system tailored to transportation

choices. This system analyzes passengers’ historical

preferences and usage patterns to identify similarities

among them.

To recommend the transportation mode most tai-

lored to a new passenger proﬁle, we use the k-Nearest

Neighbors (k-NN) classiﬁcation algorithm. k-NN is

a versatile and intuitive machine learning algorithm

for classiﬁcation and regression tasks. Its fundamen-

tal principle is leveraging the proximity of data points

in the feature space to make predictions for new, un-

seen instances (Airen and Agrawal, 2022). In what

follows, we describe the K-NN algorithm.

1. Input: Training data X

train

, labels y

train

, test data

test

, number of neighbors k

2. Output: Predicted labels for test data

3. For each test instance x

test

∈ X

test

(a) Compute distances between x

test

and all train-

ing instances in X

train

(b) Select the top k neighbors based on distances

among the k neighbors

Preparing and optimizing the k-NN classiﬁcation

model for predicting transportation modes is a crucial

phase. The dataset is strategically split into training

and test sets, with 80% of instances designated for

training and the remaining 20% for testing. This en-

sures a robust evaluation of the model’s performance

on unseen data.

We conducted several experiments to determine

the optimal settings for the k-NN model. These ex-

periments involved the selection of the most suitable

distance metric, identifying prominent attributes, de-

termining an appropriate k value, and assessing how

the dataset’s size inﬂuences the model’s performance.

3.2.1 Selecting the Most Appropriate Distance

Metric

In Figure 2, we present the k-NN model’s accuracy

evaluation when employing various distance met-

rics, namely Euclidean, Manhattan, Chebyshev, and

Minkowski (Nayak et al., 2022). The test accura-

cies corresponding to each distance metric are as fol-

lows: 84.78% for Euclidean, 82.61% for Manhattan,

76.09% for Chebyshev, and 84.78% for Minkowski.

For the subsequent analyses and model applications,

we will utilize the Euclidean distance, given its rel-

atively higher accuracy than other metrics. This de-

cision is based on the observed superior performance

of the Euclidean metric in capturing the underlying

patterns in our dataset. These results provide insights

into the performance of the k-NN model under dif-

ferent distance calculations. Notably, the Euclidean

and Minkowski distances exhibit similar and rela-

tively higher accuracies compared to the Manhattan

and Chebyshev distances. This suggests that, in the

context of our dataset and problem domain, the dis-

tance metric choice signiﬁcantly impacts the k-NN al-

gorithm’s predictive performance. The observed vari-

ations in accuracy underscore the importance of care-

fully selecting an appropriate distance metric tailored

to the characteristics of the data, as it can inﬂuence

the model’s ability to capture underlying patterns and

relationships.

3.2.2 The Impact of Attributes Elimination on

Model Performance

At the outset of our analysis, we conducted a series of

experiments to assess the impact of attribute elimina-

tion on the accuracy and F1-Score of our transporta-

tion mode recommendation model. This investigation

systematically removed different sets of attributes re-

lated to ’Function,’ ’Reason for travel,’ and ’Days of

travel’ from the dataset. The results of these exper-

iments were then visualized in the curve depicted in

Figure 3, where each point corresponds to a distinct

conﬁguration of attribute elimination.

The x-axis of the graph indicates the number of

eliminated attributes, while the y-axis showcases the

HybridCRS-TMS: Integrating Collaborative Recommender System and TOPSIS for Optimal Transport Mode Selection

387

Figure 2: Accuracy related to each distance metric.

associated accuracy and F1-Score values. This graph-

ical representation allows us to discern patterns and

trends related to attribute elimination and better un-

derstand its effects on the model’s performance.

Elimination of ’Days,’ ’Function,’ and ’Reason’

resulted in the highest performance across all met-

rics. This suggests that these attributes may introduce

noise or redundancy to the model, and their concur-

rent removal enhances overall predictive accuracy.

3.2.3 The Selection of Best k Neighbour Value

To optimize the K-NN model, we utilized Grid-

SearchCV, a technique that systematically searches

through a speciﬁed parameter grid to ﬁnd the com-

bination that yields the best performance.

We present in Figure 4 the validation curve for

the k-Nearest Neighbors classiﬁcation model. This

curve illustrates the relationship between the number

of neighbors (k) and the model’s performance met-

rics, speciﬁcally the cross-validation score. The x-

axis represents different values of k, while the y-axis

showcases the corresponding cross-validation scores.

The validation curve is a crucial visualization tool that

allows us to explore how changes in the hyperparam-

eter (k) inﬂuence the model’s accuracy. By examin-

ing this curve, we can identify the optimal value of

k, which is 3, often called the ”elbow” point, where

the model achieves the best balance between bias and

variance. This analysis is pivotal for making informed

decisions about hyperparameter tuning and ensuring

the robustness and generalization capability of the k-

NN model.

3.2.4 Impact of the Dataset’s Size on the Model

Performance

The learning curve illustrated in Figure 5 depicts the

model’s performance in terms of accuracy concern-

ing the size of the training set. The cross-validation

curve starts when the training set size is relatively

small x=10, and the accuracy is approximately 0.51.

As the size of the training set increases, the curve

ascends, reaching a peak at a certain point x=140

with an accuracy of around 0.8. This indicates that

adding initial training data led to an improvement in

the model’s performance. The general interpretation

of this learning curve is that the model beneﬁts from

adding more training data up to a certain point.

3.3 Recommender System’s Final

Settings and Performance

This section evaluates the ﬁnal performance of our

system based on the ﬁndings of the experiments that

were conducted. Table 2 presents the used k-NN set-

tings.

Table 2: Best Parameters for k-NN Model.

Parameter Value

Number of Neighbors (k) 3

Weighting Method Distance

Distance Metric Euclidean

The outcomes were meticulously presented and

analyzed to evaluate the efﬁcacy of our recommenda-

tion system, leveraging the k-NN classiﬁcation model

with optimal parameters: 3 neighbors and distance-

based weighting. The selection of these parameters

signiﬁes that, for the prediction, the system considers

the three nearest neighbors with distance-weighted

voting. This tailored approach ensures that our rec-

ommendation system operates with precision, taking

into account the characteristics and preferences of

users for a more personalized and effective transporta-

tion mode suggestion.

Our recommender system proves good perfor-

mance, with a cross-validation accuracy of 78.65%,

ﬁnal model test accuracy of 84.78%.

ICSOFT 2024 - 19th International Conference on Software Technologies

388

Figure 3: Impact of attributes elimination on model accuracy.

Figure 4: Validation curve.

In the context of supervised learning algorithms

and, more speciﬁcally, concerning our k-NN classiﬁ-

cation algorithm, we evaluate the precision and recall

values. These metrics are the two important ones used

to evaluate the model’s performance, especially in bi-

nary or multilabel classiﬁcation problems (Nguyen

et al., 2023). Equations 2 and 3 present the precision

and recall formulas used to evaluate our model.

Precision =

True Positive

True Positive + False Positive

(2)

Recall =

True Positive

True Positive + False Negative

(3)

The precision value of our model is 0.9600, which

means that approximately 96% of the instances pre-

dicted as positive by the model are true positives. This

is a high precision value, indicating that it is quite ac-

curate when the model predicts a positive class. Con-

cerning the recall value, which is 0.9231, it means that

the model has captured about 92.31% of all true pos-

itive instances in the dataset. Finally, our model has

an impressive F1-Score of 94.12%. We think that this

HybridCRS-TMS: Integrating Collaborative Recommender System and TOPSIS for Optimal Transport Mode Selection

389

Figure 5: Learning curve.

is a high recall value, suggesting that the model is ef-

fective at identifying a large proportion of the actual

positive cases.

3.4 An MCDM Approach for Objective

Transport Mode Selection

To optimize the selection of a transport mode, we pro-

pose a hybrid decision-making approach that com-

bines traditional objective criteria with the person-

alized recommendations of our collaborative ﬁlter-

ing recommender system. Our multi-criteria decision

method integrates factors such as environmental im-

pact, security, and duration, ensuring a comprehen-

sive evaluation. The subjective element is introduced

by treating the output of the recommender system as

a distinct criterion, capturing user preferences.

Multicriteria Decision Making (MCDM) meth-

ods allow considering multiple criteria simultane-

ously during the decision-making process. The main

objective of MCDM is to provide tools and techniques

to assess, rank, and choose among alternatives, con-

sidering several factors or criteria. These methods

are particularly useful when decisions involve mul-

tiple and often conﬂicting considerations, requiring

a comprehensive evaluation. MCDM provides sys-

tematic frameworks to structure and analyze complex

problems by integrating qualitative and quantitative

information. These methods assist in prioritizing op-

tions, evaluating trade-offs, and facilitating informed

decision-making in contexts where multiple criteria

need to be considered. In summary, MCDM aims to

aid decision-makers in navigating complex situations

by providing systematic and objective approaches to

assess and compare different alternatives (Taherdoost

and Madanchian, 2023).

To propose a comprehensive system incorporating

the subjective and objective criterion and to perform

the ﬁnal selection, we will employ the Technique for

Order of Preference by Similarity to Ideal Solution

(TOPSIS) (Chakraborty, 2022) as our multi-criteria

decision method. This approach allows for a bal-

anced consideration of both objective and subjective

aspects, facilitating a well-rounded and personalized

transport mode recommendation.

TOPSIS approach is widely employed for ranking

and selecting alternatives in decision-making scenar-

ios involving multiple criteria. TOPSIS operates in

several key steps: it starts by normalizing a decision

matrix to ensure uniformity in scale across different

criteria. If criteria have varying levels of importance,

weights can be assigned to reﬂect their relative sig-

niﬁcance. The method deﬁnes both an ideal solu-

tion, representing the best possible performance for

each criterion, and an anti-ideal solution, represent-

ing the worst performance. The Euclidean distance

or other distance measures are then used to calculate

the proximity of each alternative to these solutions.

Based on this distance calculation, TOPSIS assigns

similarity scores to alternatives. The closer an alter-

native is to the ideal solution, the farther it is from

the anti-ideal solution, the higher its similarity score.

In the ﬁnal step, alternatives are ranked according to

similarity scores, providing a clear preference order.

TOPSIS is valued for its simplicity and effectiveness

ICSOFT 2024 - 19th International Conference on Software Technologies

390

in handling both positive and negative aspects of de-

cision criteria. It offers decision-makers a straightfor-

ward method for selecting the most preferred option

in complex decision environments.

The versatility of TOPSIS is evident in its ex-

tensive application across various domains. It has

been successfully employed in diverse ﬁelds such as

purchase decisions and outsourcing provider selec-

tion (Kahraman et al., 2009), manufacturing decision-

making ﬁnancial performance analysis, service qual-

ity assessment, educational selection applications,

technology selection, material selection, product se-

lection, strategy evaluation, and critical mission plan-

ning. This broad spectrum of applications under-

scores the adaptability and effectiveness of TOPSIS in

addressing decision-making challenges across differ-

ent contexts(Chiharu Nanayakkara and Moayedikia,

2020).

3.4.1 TOPSIS Methodology

In this section, we will describe the key steps of the

TOPSIS method.

• Normalization: The ﬁrst step involves normaliz-

ing the decision matrix, denoted as X, where x

i j

represents the performance of alternative i on cri-

terion j. Normalization is typically achieved us-

ing the Min-Max normalization method.

• Weighting: If criteria have different importance

levels, weights (w

) can be assigned. The

weighted normalized decision matrix is then ob-

tained:

′

i j

= w

· x

′

i j

• Ideal and Anti-Ideal Solutions: The ideal solution

∗

) and anti-ideal solution (A

−

) are determined

based on the nature of the criterion (maximization

or minimization):

∗

= max(v

′

), A

−

= min(v

′

)

where v

′

represents the j-th column of the

weighted normalized decision matrix.

• Distance calculation: The Euclidean distance (D

and D

−

) is then computed for each alternative

concerning the ideal and anti-ideal solutions:

∑

j=1

′

i j

− A

∗

)

, D

−

∑

j=1

′

i j

− A

−

)

where m is the number of criteria.

• Similarity Scores: The relative closeness of each

alternative is assessed using the following similar-

ity score:

−

+ D

−

• Ranking: Alternatives are ranked based on sim-

ilarity scores, with higher scores indicating a

higher preference.

TOPSIS is known for its simplicity and effectiveness

in handling both positive and negative aspects of de-

cision criteria. It provides a clear ranking of alter-

natives, aiding decision-makers in selecting the most

preferred option based on multiple criteria.

3.4.2 Decision Objective Criteria

Criteria used for a hybrid transport mode (Bus, shared

taxi, individual taxi, and car) selection are Recom-

mendation by similar passenger, Environmental im-

pact, Security, and Duration

• Recommendation by a similar passenger: Binary

(0 for not recommended and 1 for recommended).

The ”recommendation by similar passenger cri-

terion” introduces a subjective element into the

decision-making process by considering the out-

put of our recommender system as a distinct cri-

terion. This criterion captures user preferences by

assigning a binary score, where 1 indicates that

the transportation mode is recommended by the

collaborative ﬁltering recommender system, and

0 denotes a non-recommended mode. In the ﬁnal

decision-making process, this criterion holds sig-

niﬁcance as it encourages the selection of a trans-

port mode recommended by our collaborative ﬁl-

tering recommender system. This approach aligns

the decision with user preferences inferred from

similar passengers, aiming to enhance the trav-

eler’s overall satisfaction and personalized expe-

rience.

• The environmental impact of transportation

modes, speciﬁcally CO

emissions, is a crucial

criterion for objective mode selection. The emis-

sions calculation considers various factors, in-

cluding distance traveled, fuel efﬁciency, and

emission factors. For public transportation modes

like buses and taxis, the emissions formula incor-

porates a Passenger Occupancy Factor (POF) to

account for the inﬂuence of passenger numbers:

Emissions

bus/taxi

=Distance × Fuel Efﬁciency

× Emission Factor × POF

(4)

HybridCRS-TMS: Integrating Collaborative Recommender System and TOPSIS for Optimal Transport Mode Selection

391

In this formula, POF indicates the ratio of passen-

gers on board relative to the vehicle’s maximum

capacity. Higher POF values result in reduced

emissions per passenger.

For individual modes such as cars or individual

taxis, the occupancy factor is usually ﬁxed at 1, 2

respectively (assuming one passenger per car and

2 passengers per individual taxi). In these cases,

the impact of occupancy on emissions is less sig-

niﬁcant, and the formula simpliﬁes to:

Emissions

car/taxi

=Distance × Fuel Efﬁciency

× Emission Factor

(5)

The Distance metric in the formula refers to the

anticipated distance that the new passenger will

travel based on their indicated departure and ar-

rival stops. This information is derived from route

planning algorithms that estimate the distance be-

tween two speciﬁed locations. It signiﬁes the spa-

tial extent that the passenger is expected to cover

during their intended journey. This personalized

distance parameter ensures that the emissions cal-

culation is tailored to the unique travel require-

ments of each passenger, contributing to a more

precise estimation of the environmental impact.

The Fuel Efﬁciency metric (see equation 6) is a

crucial component in the calculation of carbon

dioxide CO

emissions for different transporta-

tion modes. It reﬂects the efﬁciency of a vehicle

in utilizing fuel to generate the required energy

for propulsion. Generally measured in units like

Miles Per Gallon (MPG) for traditional vehicles

or equivalent metrics for alternative fuel sources,

the fuel efﬁciency value indicates how far a ve-

hicle can travel on a speciﬁc amount of fuel. In

the context of our environmental impact assess-

ment, higher fuel efﬁciency values are desirable

as they denote a more eco-friendly performance

with fewer emissions produced per unit of dis-

tance traveled. This metric plays a signiﬁcant role

in evaluating the sustainability of each transporta-

tion mode, aligning with the broader goal of min-

imizing the carbon footprint associated with pas-

senger journeys.

Fuel Efﬁciency =

Distance

Fuel Consumption

(6)

Distance represents the total distance traveled by

the vehicle, which can be the distance between

a speciﬁc passenger’s departure and arrival stops.

Fuel Consumption is the vehicle’s fuel consump-

tion during the journey. This value is speciﬁc

to each mode of transportation and can be ob-

tained from vehicle speciﬁcations or real-world

measurements.

• Security: Accident statistics play a crucial role in

assessing the security level of each mode of trans-

portation. The security level is determined by ana-

lyzing historical accident data, providing valuable

insights into the safety performance of different

transportation modes. A lower accident rate is in-

dicative of a higher security level. The formula

used for computing the security level is expressed

as:

Security Level = 1 −

Accident Rate

Max Accident Rate

(7)

In this formula, Accident Rate represents the his-

torical accident rate speciﬁc to each mode of

transportation, while Max Accident Rate is a hy-

pothetical maximum accident rate. The resulting

Security Level is a normalized value between 0

and 1, where 0 indicates a lower security level

(higher accident rate), and 1 signiﬁes a higher se-

curity level (lower accident rate). It’s essential to

note that the accident data utilized in this evalua-

tion is sourced from the Tunisian National Road

Safety Observatory (ONSR). For detailed acci-

dent statistics, the interested reader can refer to

the ONSR website

• Duration: Estimating the average duration for

each mode of transportation involves using the

TrackGPS tool

. This tool provides valuable

insights into the average duration of traveling

routes, considering the habitual paths of each

transportation mode. Additionally, it factors in the

average duration of stops for public transportation

modes. By leveraging the capabilities of Track-

GPS, we can obtain reliable and real-world data to

enhance the accuracy of our duration assessments.

3.4.3 Use Case of TOPSIS Evaluation

In our study, we sought the expertise of a transport

specialist to assign appropriate weights to various cri-

teria. The expert, utilizing a ten-point scale where

1 indicates not at all important criterion and 10 de-

notes very important criterion, meticulously evaluated

transportation criteria (see Table 4.

In our evaluation of transportation alternatives,

various criteria were considered to provide a com-

prehensive assessment of each mode. Table 3 sum-

marizes the data evaluation for four transportation

https://onsr.nat.tn/onsr/index.php?page=3fr

https://trackgps.ro/en/

ICSOFT 2024 - 19th International Conference on Software Technologies

392

Table 3: Data Evaluation for Transportation Alternatives.

Alternative Recommendation by similar passengers Environmental Impact Security Duration

Car 1 1.68 6.8 20

Shared Taxi 0 7.29 7.2 30

Individual Taxi 1 2.92 4.5 25

Bus 1 0.192 9.0 35

Table 4: Weights assigned by the transport specialist to eval-

uation criteria.

Criteria Weight

Recommendation by similar passengers 6

Environmental Impact (CO

Emissions) 8

Security (Accident Statistics) 8

Duration 9

alternatives: Car, Shared Taxi, Individual Taxi, and

Bus. The ”Recommendation by Similar Passengers”

column reﬂects a binary value, indicating whether

the transportation mode is recommended by similar

passengers (1 for recommended, 0 for not recom-

mended).

The ”Environmental Impact (kg CO

)” column

represents the estimated CO

emissions for each

transportation mode over a distance of 7 km, assum-

ing the use of gasoil as fuel. Notably, the values range

from 0.192 kg CO

for the Bus, known for its eco-

friendly features, to 7.29 kg CO

for the Shared Taxi,

reﬂecting its potentially higher environmental impact.

The ”Security” column assigns security scores

to each mode, considering factors such as histori-

cal accident statistics. Higher security scores indi-

cate modes with lower accident rates, contributing to

a safer travel experience. For instance, the Bus re-

ceived a security score of 9.0, emphasizing its per-

ceived safety.

The ”Duration” column provides the duration for

each mode, representing the average time it takes to

complete the 7-kilometer journey. These durations,

measured in minutes, were estimated using Track-

GPS, a tool that tracks and estimates transportation

durations based on real-world data. The ﬁnal re-

sults corresponding to the TOPSIS ranking and scores

are presented in Table 5. The TOPSIS analysis pro-

vides valuable insights into the performance of differ-

Table 5: TOPSIS scores and ranking for transportation al-

ternatives.

Alternative TOPSIS Score Rank

Car 0.45812809 2

Shared Taxi 0.41078759 3

Individual Taxi 0.39101742 4

Bus 0.66696562 1

ent transportation alternatives based on multiple crite-

ria, including recommendation by similar passengers,

environmental impact (CO

emissions), security (ac-

cident statistics), and duration. The results reveal a

comprehensive evaluation, with each alternative as-

signed a TOPSIS score and corresponding rank.

Starting with the Bus, it emerges as the top-

ranking alternative, securing the lowest TOPSIS score

of 0.3910. This indicates that the Bus performs excep-

tionally well across the considered criteria, showcas-

ing the most favorable balance and proximity to the

ideal solution.

The Car follows closely with a TOPSIS score of

0.4108, earning the second position in the ranking.

While it performs well, it falls just short of the Bus

alternative in achieving an optimal balance across the

criteria.

The Shared Taxi takes the third position with a

TOPSIS score of 0.4581. It demonstrates a good over-

all performance but is outranked by both the Bus and

Car alternatives.

Finally, the Individual Taxi secures the highest

TOPSIS score of 0.66697 but ranks fourth. Although

it excels in certain criteria, the overall evaluation

places it behind the other alternatives.

In summary, the TOPSIS analysis suggests that,

for a speciﬁc new passenger proﬁle with predeﬁned

preferences, the Bus stands out as the most favorable

transportation mode, offering a well-balanced perfor-

mance across various criteria. At the same time, the

Individual Taxi, despite being a good choice in certain

aspects, falls behind in the overall ranking.

4 CONCLUSION

This paper presented a comprehensive two-phased

approach for transportation mode recommendation,

blending collaborative ﬁltering with the k-NN al-

gorithm and the TOPSIS method. The initial

phase focuses on reﬁning subjective recommenda-

tions through collaborative ﬁltering, ensuring accu-

rate and personalized suggestions for users. The sub-

sequent phase employs TOPSIS to introduce an ob-

jective dimension, evaluating modes based on crite-

ria such as environmental impact, security, and dura-

tion. This hybrid decision-making model combines

HybridCRS-TMS: Integrating Collaborative Recommender System and TOPSIS for Optimal Transport Mode Selection

393

both subjective user preferences and objective evalu-

ations, offering a more nuanced and comprehensive

solution to transportation mode selection.

Furthermore, the proposed approach stands out

for its ability to integrate both objective and sub-

jective factors seamlessly. By incorporating user

preferences and incorporating objective criteria, the

hybrid decision-making model aims to provide a

well-rounded recommendation system. This unique

combination enhances the robustness and adaptabil-

ity of the system, catering to individual user needs

while considering broader performance indicators.

As a result, the hybrid model introduces a balanced

and effective approach to transportation mode selec-

tion, fostering a more sustainable and user-centric

decision-making process.

As future work, we envision a further investiga-

tion that involves expanding our research in two key

areas: collecting more comprehensive passenger data

and evaluating additional recommender system algo-

rithms.

REFERENCES

Airen, S. and Agrawal, J. (2022). Movie recommender

system using k-nearest neighbors variants. National

Academy Science Letters, 45(1):75–82.

Alhijawi, B. and Kilani, Y. (2020). A collaborative ﬁltering

recommender system using genetic algorithm. Infor-

mation Processing & Management, 57(6):102310.

Arnaoutaki, K., Bothos, E., Magoutas, B., Aba, A., Es-

zterg

ar-Kiss, D., and Mentzas, G. (2021). A recom-

mender system for mobility-as-a-service plans selec-

tion. Sustainability, 13(15).

Arnaoutaki, K., Magoutas, B., Bothos, E., and Mentzas, G.

(2019). A hybrid knowledge-based recommender for

mobility-as-a-service. In ICETE (1), pages 101–109.

Chakraborty, S. (2022). Topsis and modiﬁed topsis: A

comparative analysis. Decision Analytics Journal,

2:100021.

Chiharu Nanayakkara, William Yeoh, A. L. and

Moayedikia, A. (2020). Deciding discipline,

course and university through topsis. Studies in

Higher Education, 45(12):2497–2512.

Hamadneh, J. and Eszterg

ar-Kiss, D. (2023). The prefer-

ences of transport mode of certain travelers in the age

of autonomous vehicle. Journal of Urban Mobility,

3:100054.

Kahraman, C., Engin, O.,

Ozg

ur Kabak, and

Ihsan Kaya

(2009). Information systems outsourcing decisions

using a group decision-making approach. Engineering

Applications of Artiﬁcial Intelligence, 22(6):832–841.

Artiﬁcial Intelligence Techniques for Supply Chain

Management.

Kedar Potdar, Taher S. Pardawala, C. D. P. (2017). A com-

parative study of categorical variable encoding tech-

niques for neural network classiﬁers. International

Journal of Computer Applications, 175(4):7–9.

Lai, Z., Wang, J., Zheng, J., Ding, Y., Wang, C., and Zhang,

H. (2023). Travel mode choice prediction based on

personalized recommendation model. IET Intelligent

Transport Systems, 17(4):667–677.

Liu, Y., Lyu, C., Liu, Z., and Cao, J. (2021). Exploring a

large-scale multi-modal transportation recommenda-

tion system. Transportation Research Part C: Emerg-

ing Technologies, 126:103070.

Mariscal, G., Marb

an, O., and Fern

andez, C. (2010). A sur-

vey of data mining and knowledge discovery process

models and methodologies. The Knowledge Engineer-

ing Review, 25(2):137–166.

Nayak, S., Bhat, M., Reddy, N. V. S., and Rao, B. A. (2022).

Study of distance metrics on k - nearest neighbor al-

gorithm for star categorization. Journal of Physics:

Conference Series, 2161(1):012004.

Nguyen, L. V., Vo, Q.-T., and Nguyen, T.-H. (2023). Adap-

tive knn-based extended collaborative ﬁltering recom-

mendation services. Big Data and Cognitive Comput-

ing, 7(2).

Rodriguez-Valencia, A., Ortiz-Ramirez, H. A., Simancas,

W., and Vallejo-Borda, J. A. (2022). Understanding

transit user satisfaction with an integrated bus system.

Journal of Public Transportation, 24:100037.

Sun, X. and Wandelt, S. (2021). Transportation mode

choice behavior with recommender systems: A case

study on beijing. Transportation Research Interdisci-

plinary Perspectives, 11:100408.

Taherdoost, H. and Madanchian, M. (2023). Multi-criteria

decision making (mcdm) methods and concepts. En-

cyclopedia, 3(1):77–87.

Wu, F., Lyu, C., and Liu, Y. (2022). A personalized rec-

ommendation system for multi-modal transportation

systems. Multimodal Transportation, 1(2):100016.

ICSOFT 2024 - 19th International Conference on Software Technologies

394