https://doi.or g/10.31449/inf.v48i6.5234 Informatica 48 (2024) 1 17–130 1 17
Machine Learning Algorithms for T ransportation Mode Pr ediction: A
Comparative Analysis
Samer Murrar
1
, Fatima Alhaj
1,∗ and Mahmoud H. Qutqut
2
1
Faculty of Information T echnology , Applied Science Private University , Amman, Jordan
2
Faculty of Computer Science, University of New Brunswick, Fredericton, NB, E3B 5A3 Canada
E-mail: 202215019@students.asu.edu.jo, f_alhaj@asu.edu.jo, m.qutqut@unb.ca
*Corresponding author
Keywords: transportation mode, transportation mode prediction, travel mode identifier , GPS trajectories, machine learn-
ing (ML)
Received: September 26, 2023
This study investigates the performance of various machine learning (ML) algorithms in pr edicting trans-
portation modes fr om lar ge datasets. The investigated algorithms include Multilayer Per ceptr on (MLP),
K-Near est Neighbors (KNN), Decision T r ee (DT), Long Short-T erm Memory (LSTM), Recurr ent Neural
Network (RNN), and Logistic Regr ession. W e rigor ously evaluated each algorithm’ s performance using a
r obust set of metrics such as pr ecision, r ecall, and F1-scor e. This study compr ehensively explains the al-
gorithm’ s capabilities, str engths, and potential weaknesses acr oss seven transportation categories: ’walk’,
’bike’, ’bus’, ’car ’, ’ taxi’, ’ train’, and ’ subway’. The DT model consistently outperformed the others,
demonstrating superior accuracy and an adequate balance of pr ecision and r ecall acr oss all modes of
transportation. Specifically , it achieved pr ecision, r ecall, and F1 scor es of ar ound 83% to 94% acr oss all
categories. These findings underline the suitability of the DT model for this classification task and its po-
tential for further applications in transportation mode pr ediction based on lar ge datasets. However , other
algorithms like LSTM and RNN also showed pr omising r esults in certain categories, suggesting the value
of continued exploration of differ ent models depending on specific use cases.
Povzetek: Raziskava pr eučuje učinkovitost algoritmov str ojnega učenja pri napovedovanju načinov pr e-
voza iz obsežnih podatkovnih zbirk, pri čemer izstopa model odločitvenih dr eves.
1 Intr oduction
The complexities of how people move within a community
- their travel behaviors and transportation choices - play a
critical role in many aspects of urban planning and devel-
opment [ 1 ]. This intricate mosaic of movement patterns
is a valuable tool for policymakers, transportation plan-
ners, and urban developers. It helps to predict future trans-
portation needs accurately , guides critical decision-making
processes, and promotes environmentally friendly practices
[ 24 ]. Insights gleaned from this data are used by transporta-
tion planners and policymakers to accurately forecast future
demand for various modes of transportation [ 23 ]. It pro-
vides recommendations, aiding informed decisions in in-
frastructure and service investment decisions. For exam-
ple, suppose analysis shows that a sizable proportion of the
population relies on public transport. In that case, a clear
justification exists for investments in expanding bus lines
or adding tube stations [ 3 ].
Furthermore, data on travel behavior is a valuable tool
for promoting environmentally sustainable transportation
practices. If, for example, a sizable proportion of the pop-
ulation relies solely on private automobiles for commut-
ing, this may indicate a need for more environmentally
friendly transportation options. Cycling lanes, carpooling
programs, and better public transportation are all potential
solutions [ 4 ]. Understanding travel behavior can also reveal
implications for health and safety . Assume that many peo-
ple prefer cycling but that there are many traf fic accidents
involving cyclists in the area. In that case, this troubling
trend may indicate the need for improved bike infrastruc-
ture or increased safety education [ 5 ].
Moreover , this comprehension can shed light on poten-
tial equity issues when lower -income people rely heavily
on public transportation, and the service is either inade-
quate or unaf fordable. Hence, policy changes are needed
to ensure equitable transportation access [ 2 ]. Understand-
ing travel behavior significantly impacts economic devel-
opment when deciding where to locate. Businesses in var -
ious industries, including retail, food, and entertainment,
frequently consider potential customers’ modes of trans-
portation [ 4 ]. Another critical application is for commu-
nities to understand how their populations travel t o prepare
for and respond to a disaster ef fectively . This information
can be used to predict which roads may require immedi-
ate clearance and which modes of transportation should be
restored as soon as possible [ 6 ].

1 18 Informatica 48 (2024) 1 17–130 S. Murrar et al.
Our paper aims to assess the accuracy of dif ferent ma-
chine learning (ML) algorithms for predicting transporta-
tion modes using lar ge datasets. The investigated algo-
rithms include Multilayer Perceptron (MLP), K-Nearest
Neighbors (KNN), Decision T ree (DT), Long Short-T erm
Memory (LSTM), Recurrent Neural Network (RNN), and
Logistic Regression. The process includes a review of the
performance of each algorithm, employing a comprehen-
sive range of evaluation metrics. The research seeks to
identify the strengths and weaknesses of these algorithms
in various transportation domains. The findings are crucial
in identifying the most appropriate ML methods for pre-
dicting transportation modes. Hence, the paper provides a
well-structured guide for researchers and developers in this
domain and opens up additional applications and research
possibilities.
This paper is or ganized as follows. Section 2 explores
the background information. Section 3 thoroughly reviews
the existing literature. Section 4 provides a detailed expla-
nation of our proposed approach, which includes data pre-
processing, feature construction, and the intricate aspects
of the model architecture. Section 5 thoroughly examines
our experimental findings. Finally , we wrap up our discus-
sion and d raw meaningful conclusions in Sections 6 and 7 ,
respectively .
2 Backgr ound
Understanding and predicting travel behavior is complex,
requiring using numerous data types and sophisticated an-
alytical techniques [ 8 ]. As location-acquisition technol-
ogy has advanced, GPS trajectory data has become one of
the most important sources of information for researching
human mobility patterns. By providing extensive records
of individuals’ spatial-temporal travels, these data provide
significant insights into how , why , and where people travel.
However , due to the inherent complexity and variety of hu-
man movement, extracting meaningful insights from raw
GPS trajectory data is dif ficult.
V arious computational strategies have been developed
over the years to deal with this dif ficulty . Among these,
ML algorithms have emer ged as particularly promising [ 9 ].
They can learn complex patterns from massive amounts
of data, making them ideal for jobs such as transportation
mode prediction. Decision trees, for example, have been
widely used due to their interpretability and versatility .
However , the performance of these algorithms is heav-
ily reliant on the quality of the incoming data and how it is
handled. As a result, data preprocessing and feature extrac-
tion are critical steps in model development. Data cleaning,
normalization, and encoding are frequently used to convert
raw GPS data into a format suitable for ML algorithms.
The study intends to use these approaches, specifically the
DT algorithm, to forecast transportation modes from GPS
trajectory data. This study contributes to the lar ger field
of travel behavior analysis and provides legislators, trans-
portation planners, and urban developers practical insights
[ 9 ].
3 Related work
Over the years, numerous studies have been conducted to
unravel the complexities of travel behavior and transporta-
tion mode prediction. These investigations have shed light
on various aspects of travel behavior , influencing the evo-
lution of prediction models and methodologies. Previous
research emphasizes the significance of decoding human
mobility patterns - a complex web of numerous factors in-
fluencing travel choices. These studies used a variety of
methodologies to unravel this complex issue, ranging from
traditional statistical methods to advanced ML algorithms.
Convolutional Neural Networks (CNNs) have been
widely used among these because of their ability to learn
and extract features from spatial data automatically . Re-
gardless of their advantages, CNNs require a lar ge amount
of training data for optimal performance and can be com-
putationally intensive, making them slower to train [ 10 ].
Other works were based on deep Neural Networks (DNN),
such as [ 1 1 ], [ 12 ]. Even though They are ef fective at learn-
ing and remembering long sequences, they are computa-
tionally demanding and may be prone to overfitting due to
their complexity . On the other hand, long-term Recurrent
Convolutional Network (LRCN) combines the strengths
of CNN and RNN rather than specifically incorporating
LSTMs. LRCN is intended to ef ficiently process sequential
data with spatial features by combining the spatial feature
extraction capabilities of CNNs with the temporal modeling
capabilities of RNN [ 13 ].
Other techniques were used, such as the Spatial-
T emporal Pattern Chain Network (STPC-Net), which mod-
els complex spatial-temporal patterns specifically designed
for transportation mode identification. Despite its ef ficacy ,
the model may be overly complicated for tasks where sim-
pler models would suf fice [ 14 ]. The Contrast-Enhanced
Robust Conditional Random Field (CE-RCRF) method
combines the advantages of both the Contrast Enhancement
(CE) and the Robust Conditional Random Field (RCRF)
methods. It is more complex and computationally demand-
ing than other methods for dealing with noise and uncer -
tainties in GPS data [ 1 ]. Investigating these techniques and
their ef fectiveness in predicting travel behavior has yielded
valuable insights for future research in this field. In T a-
ble 1 , the results and key findings of the numerous stud-
ies on transportation mode prediction and their respective
methodologies and performance metrics, including the F1-
score, are presented in detail. Our approach can potentially
provide significant insights into this multifaceted area de-
spite its simplicity .

Machine Learning Algorithms for T ransportation Mode… Informatica 48 (2024) 1 17–130 1 19
T able 1: Summary of Related W ork on T ransportation Mode Prediction
Study Resear ch Focus Methodologies F1-Scor e % Findings
[ 10 ] Feature extraction from
spatial data
Convolutional Neural
Networks (CNNs)
81.77 Ef fective but require lar ge data and
are computationally intensive
[ 1 1 ] T ravel Mode Identifica-
tion using GPS
W avelet T ransform and
Deep Learning
80.53 Ef fective at learning long se-
quences, computationally demand-
ing
[ 12 ] GPS T rajectory analysis Semi-Supervised Feder -
ated Learning
80.83 Ef ficient for sequential data, prone
to overfitting
[ 13 ] GPS data-based mobility
mode inference
Long-term Recurrent
Convolutional Networks
(LRCNs)
82.30 Combines CNN and RNN, suitable
for sequential spatial data
[ 14 ] Spatial-temporal data
analysis
Spatio-T emporal Point
Clouds (STPC-Net)
81.05 Ef fective for complex patterns, may
be overly complex for simpler tasks
[ 1 ] T ravel mode identifica-
tion from GPS tracks
Sequence-to-sequence
model, Deep Learning
85.41 Highly accurate, ef fective for com-
plex GPS data
4 Pr oposed methodology
Our paper proposes a comprehensive framework for T ravel
Mode Identification that includes four steps: namely , data
preprocessing, feature construction, predictive models, and
evaluation methods. These modules collaborate to form a
unified pipeline for identifying travel modes ef fectively and
ef ficiently .
The first step, data preprocessing, is crucial in preparing
raw data for further analysis. This step involves cleaning
and standardizing it to ensure the quality and consistency
of the data. W e provided the reliability of the data used
for travel mode identification by addressing missing val-
ues, outliers, and inconsistencies. This step also included
extracting meaningful features. These characteristics are
chosen based on their relevance and potential impact on
travel mode identification.
Then, we employed various predictive models to clas-
sify travel modes based on the constructed features. Our
goal was to juxtapose the performance of these models to
discern the most ef fective one(s) for our specific task. The
models used include Multilayer Perceptron (MLP), Long
Short-T erm Memory (LSTM), Recurrent Neural Network
(RNN), Decision T rees, Logistic Regression, and K-nearest
Neighbors (KNN) algorithms. Each model was trained on
the same training data set and evaluated on a standard test-
ing set to ensure a fair comparison. The specific configura-
tions and hyperparameters selected for each ML model play
a critical role in determining the accuracy of our research
findings. These settings are essential in optimizing each
model’ s performance and ensuring our results’ validity . In
MLP configuration, the model comprised three layers with
64, 128, and 256 neurons, respectively . ReLU activation
functions were utilized. The learning rate was set at 0.001,
and the model was trained for 100 epochs.
For LSTM networks, our model included 50 LSTM units,
incorporating a dropout rate of 0.2 to prevent overfitting.
A learning rate of 0.001 was maintained during the training
phase. A similar approach was adopted for RNN, wherein
the model comprised 50 RNN units with a dropout rate of
0.2. The training process was conducted using a learning
rate identical to the LSTM model’ s. The Decision T ree
(DT) model was structured with a maximum depth of 10,
utilizing the Gini index as the criterion for data splitting.
In the Logistic Regression model, an L2 regularization ap-
proach was implemented with a regularization strength (C)
1.0.
For the KNN algorithm, we selected a configuration of
five neighbors, balancing computational ef ficiency with
prediction accuracy . The configurations and hyperparame-
ters for each model were meticulously determined through
a combination of grid search and empirical testing. This
approach was undertaken to optimize performance on our
validation data set. The judicious selection of these param-
eters is crucial in shaping the model’ s ability to ef fectively
learn from the training data and generalize to new , unseen
data. This process ensures that our models are well-tuned
for the task at hand and robust in their application to diverse
data scenarios.
In the final step, we evaluate the ef ficacy of the various
predictive models that we have used. W e use a set of perfor -
mance metrics for this purpose, including accuracy , preci-
sion, recall, and the F1 score. These metrics enable us to as-
sess each model’ s ability to identify travel modes correctly .
Every individual model is subjected to a thorough evalua-
tion, providing us with detailed insights into the model’ s
strengths, weaknesses, and distinguishing characteristics.
In evaluating our models, we analyzed the ef fects of the
selected configurations and hyperparameters on important
metrics, including accuracy , precision, recall, and the F1
score. The precision and recall scores of the DT and KNN
models were significantly af fected by the depth of the De-
cision T ree and the number of neighbors in the KNN. This
emphasizes the need for careful hyperparameter tuning.
In this study , we use the extensive geographic data con-
tained in Microsoft’ s GeoLife GPS T rajectory 1.3 dataset

120 Informatica 48 (2024) 1 17–130 S. Murrar et al.
1
, a robust repository that includes a wealth of information
about human mobility patterns [ 7 ]. The GeoLife GPS T ra-
jectory 1.3 dataset from Microsoft is a rich repository of
geographic data that provides a comprehensive view of mo-
bility patterns, making it an invaluable resource for geospa-
tial researchers and developers. This dataset, derived from
various location-enabled devices, enables a thorough ex-
amination of spatial and temporal behaviors [ 7 ]. The GPS
T rajectory 1.3 dataset, created as part of Microsoft’ s Geo-
Life project, contains the mobility data of 182 users from
April 2007 to August 2012. This massive dataset includes
17,621 trajectories and over 24.7 million individual loca-
tion points [ 7 ].
Each trajectory in the dataset is a series of timestamped
points that provide location information and a chronologi-
cal perspective necessary for understanding movement pat-
terns over time [ 7 ]. The location points were recorded at
five-second intervals, resulting in a high-resolution view of
each trajectory . This geographically diverse dataset cov-
ers a wide range of areas in over 30 Chinese cities. Be-
cause of the broad geographic scope, comparative studies
of mobility patterns in various cultural and urban contexts
are possible [ 7 ]. Furthermore, one distinguishing feature
of this dataset is that it includes a wealth of associated in-
formation in addition to geographic and temporal data. For
some users, a mode of transportation is available, provid-
ing insight into the options of walking, cycling, driving, or
taking public transportation. This extra data layer can be in-
strumental in studies examining transportation options and
travel behavior [ 7 ].
4.1 Data pr epr ocessing
Data preprocessing was the first step in preparing our
dataset for further research. This stage ensured that the
dataset’ s format was standardized, that unnecessary at-
tributes were removed, and that all necessary changes were
made. The following preprocessing procedures were car -
ried out:
1. Data Integration:
Data from 18,670 files belonging to 182 individuals
were combined into a single data file, similarly inte-
grating trajectory labels from 69 users. After export-
ing the trajectory data points to a unified dataset, they
were linked with their labels, yielding approximately
24,876,978 records.
2. Data Reduction:
The data reduction process entailed removing irrele-
vant attributes from the dataset to streamline it. This
included removing the ’Param’ attribute, which held
no informational value as it was consistently zero
across all instances. Furthermore, due to the preva-
lence of undefined and inconsistent values, the ”Al-
titude” attribute was also eliminated. Finally , cases
1
https://www .microsoft.com/en-us/research/publication/geolife-gps-
trajectory-dataset-user -guide/
lacking labels and those with a zero value for the
time attribute were systematically eliminated from the
dataset to ensure data integrity and assist the super -
vised learning requirement.
4.2 Featur e construction
Following data preprocessing, the subsequent crucial stage
is feature construction, aiming to establish meaningful fea-
tures as valuable inputs for the modeling process. The fol-
lowing steps were taken to complete this process:
1. Attribute Generation:
The process of the feature creation procedure began
by extracting critical attributes from the existing GPS
coordinates and timestamps. T o begin, the property
denoting the distance to the following location was de-
termined in kilometers using the equation ( 1 ). Using
the equation ( 2 ), the time to the following location was
then calculated in hours. The velocity attribute was in-
troduced, which was calculated as the distance-to-time
ratio and expressed in kilometers per hour using the
equation ( 3 ). This change improved the dataset’ s ac-
ceptability for further analysis and added an essential
predictive feature to the model.
In addition to these primary attributes, the dataset was
further enhanced by computing two more nuanced at-
tributes. These included the acceleration, which was
stated in kilometers per hour squared and calculated
using equation ( 4 ), and the angular velocity , which
was expressed in radians per hour and calculated us-
ing equation ( 5 ). Including these complex features in-
creased the dataset’ s analytical reach, of fering more
profound and detailed insights for the following stages
of analysis and modeling.
2. Outlier Removal
T o ensure data integrity and enhance our analysis’ s ro-
bustness, we eliminated any dataset instances display-
ing dubious or physically impossible travel situations.
Cases with negative V elocity , T ime to the next point,
or Distance to the following point values, in particu-
lar , were immediately eliminated. This critical phase
aided in the removal of data irregularities and other er -
rors that may have occurred during the data collection
procedure.
3. Data Constraint Application
W e established speed limits for each distinct mode of
transportation after removing these outliers. This re-
quired setting average speed limits for various trans-
portation modes, including walking, biking, and other
motorized and public transportation types.
Instances that exceeded the established speed limits
were deemed abnormal and were removed from the
dataset. Our dataset remained grounded in reason-
able travel conditions by adhering to realistic speed

Machine Learning Algorithms for T ransportation Mode… Informatica 48 (2024) 1 17–130 121
Figure 1: Formula for calculating the distance between two points
d=2× R× arcsin
( √ sin
2
( Lat2− Lat1
2
) + cos( Lat1)× cos( Lat2)× sin
2
( Long2− Long1
2
) ) (1)
where:
-d represents the distance to the next point in kilometers (km),
-2 represents a constant factor used in the calculation,
-R is the radius of the Earth in kilometers, taken to be approximately 6,371 km,
-Lat1 andLat2 are the latitude coordinates of the two points,
-Long1 andLong2 are the longitude coordinates of the two points.
Figure 2: Formula for calculating the time to reach the next point
t=( Datetime2− Datetime1)× 24.0 (2)
where:
-t represents the time duration between two points in hours (h),
-Datetime1 andDatetime2 are the timestamps of the two points.
Figure 3: Formula for calculating the velocity
v =
d
t
(3)
where:
-v represents the velocity in kilometers per hour (km/h),
-d represents the distance to the next point in kilometers,
-t is the time taken to travel from the first point to the second point in hours.
Figure 4: Formula for calculating the acceleration
a=
v
2
− v
1
t
(4)
where:
-a represents the acceleration in kilometers per hour squared (km/h
2
),
-v
1
andv
2
are the initial and final velocities respectively ,
-t represents the time interval.
Figure 5: Formula for calculating the angular velocity
∆ T erm= sin
2
( Lat2− Lat1
2
) + cos( Lat1)× cos( Lat2)× sin
2
( Long2− Long1
2
) (5)
av =
2× atan2
( √ ∆ T erm, √ 1− ∆ T erm
) R× t
(6)
where:
-av represents the angular velocity in radians per hour (degrees/h),
-Lat1 andLat2 are the initial and final latitudes, respectively , in radians,
-Long1 andLong2 are the initial and final longitudes, respectively , in radians,
-R is the radius of the Earth, taken to be approximately 6,371 km,
-t represents the time interval.

122 Informatica 48 (2024) 1 17–130 S. Murrar et al.
T able 2: Speed Limits for Each Mode of T ravel
T ravel Mode Speed Limit (km/h)
W alk 12
Bike 50
Car 160
T axi 140
Bus 120
Subway 150
T rain 320
T able 3: Class Distribution After Data Preprocessing and
Feature Construction.
T ransportation
Mode
Counts Per centage (%)
W alk 1,497,710 28.16
Bus 1,275,389 23.98
Bike 945,077 17.77
T rain 560,528 10.54
Car 51 1,585 9.62
Subway 286,1 12 5.38
T axi 241,976 4.55
constraints. The speed limits for each mode of trans-
portation are depicted in T able 2 . The study’ s reliabil-
ity and accuracy were significantly improved by this
thorough approach to feature development, which in-
cluded screening out rare situations and adhering to
strict travel mode speed thresholds.
4.3 T ravel mode identifier
The cornerstone of every ML endeavor is u ndeniably the
dataset in use. As we transition from the data prepro-
cessing and feature construction phases into model devel-
opment, it becomes imperative to understand the refined
dataset. The characteristics of this dataset illuminate the
intricacies of the problem at hand and hint at potential chal-
lenges that might arise during model construction. Post-
processing, our dataset encompasses 5,318,377 instances,
each assigned to one of seven distinct classes. T able 3
shows a detailed breakdown of these classes.
Our approach used MLP , Logistic Regression, KNN,
DT , LSTM, and RNN to compare and assess various classi-
fiers’ performance. These classifiers were chosen for their
ability to handle multiclass classification tasks in various
contexts. Each classifier has its implementation method,
but all share a common foundation of preprocessing steps
and performance evaluation metrics.
ML algorithms were implemented using Python in
Google Colab
2
. The initial steps for all algorithms were
similar . W e imported the necessary libraries, loaded the
data into a pandas DataFrame, and performed preliminary
data preprocessing. Initially , the dataset was retrieved from
2
https://colab.research.google.com/
a CSV file stored on Google Drive. W e then used dic-
tionary mapping to convert categorical variables, such as
’T ransportationMode’, ’UserCode’, and ’T rajectoryCode’
into numerical form.
W e used a predefined dictionary ’mode_dict’ to trans-
form the ’T ransportationMode’ variable, which served as
the tar get variable. W e also converted ’UserCode’ and ’T ra-
jectoryCode’ into numerical codes to make the dataset suit-
able for ML models. A percentage of the dataset was cho-
sen for subsequent analysis to ensure ef ficient processing.
After the initial preprocessing steps, we divided the dataset
into features (X) and labels (y). The ’T ransportationMode’
column was among the labels, while the others were among
the features. W e used StandardScaler to normalize the fea-
tures to ensure they were consistent. This step required
removing the mean and scaling the features to unit vari-
ance, which is necessary for many ML estimators. W e di-
vided the data into training and testing sets using the sklearn
train_test_split function to assess the models’ performance.
The training set contained 80% of the data, while the test
set received the remaining 20%. This division allowed us
to evaluate the models’ performance on previously unseen
data, ensuring a fair evaluation.
This study used multiple ML models to address our re-
search objectives. W e employed a robust set of critical
metrics to comprehensively evaluate their performance, in-
cluding accuracy , bias, variance, precision, recall, and the
F1 score. Considering these metrics, we gained valuable
insights into the identifier ’ s ef ficacy across various modes
and ascertained its precision in predicting specific modes.
W e used the DecisionT reeClassifier from sklearn to imple-
ment the DT model, training it on our data and using it to
predict the labels of the test set. Its performance was as-
sessed by comparing predicted and actual labels. W e im-
ported the Logistic Regression classifier from sklearn for
the Logistic Regression model, followed the same process
as the DT model, and evaluated the model similarly . The
KNN model was built with Sklearn’ s K-Neighbors Classi-
fier . After fitting the model to the training data, we used
the same evaluation process to make predictions on the test
set.
W e defined and built the architecture of the MLP , LSTM,
and RNN models using T ensorFlow’ s Keras API. The MLP
model had input, hidden, and output layers, with ’relu’ as
the hidden layer activation function. The LSTM model
began with an LSTM layer , while the RNN model be-
gan with a SimpleRNN layer . All three models con-
cluded with a dense layer with ’ softmax’ as the activation
function, which is appropriate for multiclass classification
problems. MLP , LSTM, and RNN models were all built
with the ’ sparse_categorical_crossentropy’ loss function,
the ’adam’ optimizer , and ’accuracy’ as a performance met-
ric. Following training, the models were used to predict test
set labels and their performance was evaluated by compar -
ing these predictions to actual labels.
Furthermore, we considered each model’ s inherent char -
acteristics and trade-of fs when determining their applica-

Machine Learning Algorithms for T ransportation Mode… Informatica 48 (2024) 1 17–130 123
Figure 6: Performance Comparison of T ransport Modes
with MLP Algorithm.
bility to our specific problem. Decision trees, for example,
are interpretable but may struggle with complex patterns. In
contrast, MLP , LSTM, and RNN models can capture such
patterns but may require more computational resources and
time for fine-tuning.
5 Results and analysis
W e used cross-validation to evaluate the ef fectiveness of the
various classifiers, including MLP , KNN, Decision T ree,
LSTM, RNN, and Logistic Regression. The five-fold cross-
validation method was explicitly used to provide a reliable
estimate of the model’ s potential performance on unseen
data, protecting against overfitting. In addition, we inves-
tigated the bias-variance trade-of f for each model to under -
stand its robustness and generalization capabilities better .
This study used six ML algorithms to model and predict
the multiclass transportation dataset. Several metrics, in-
cluding accuracy , precision, recall, and the F1-score, were
used to evaluate and compare the performance of the mod-
els. In analyzing the results, we will look at two perspec-
tives, the first from the Model point of view and the second
from the point of view of the results of the transfer mode.
The MLP model had an overall accuracy of 68.55%.
According to the confusion matrix, the model performed
best in the ’walk’ category , correctly identifying approx-
imately 89% of instances. The ’ taxi’ and ’ subway’ cate-
gories had the lowest accuracy , with only about 8% and
32% of cases correctly identified, respectively . This indi-
cates that the model has dif ficulty distinguishing between
these categories. The details of the MLP performance are
provided in T able 4 and Figure 6 .
The overall accuracy of the KNN model was 79.29%,
which ed well in all categories, with the highest accuracy
observed in the ’ train’ category (approximately 91% of in-
stances correctly identified). Conversely , the model strug-
Figure 7: Performance Comparison of T ransport Modes
with KNN Algorithm.
Figure 8: Performance Comparison of T ransport Modes
with DT Algorithm.
gled with the ’ taxi’ and ’ subway’ categories, as evidenced
by lower recall rates of 50%
The DT model outperformed the previous two algorithms
with an overall accuracy of 87.41%. It performed excep-
tionally well in distinguishing the ’ train’ category , correctly
identifying approximately 96% of instances. Interestingly ,
this model performed relatively well in the ’ taxi’ category
(approximately 83% The details of the DT model perfor -
mance, including precision and recall for each transporta-
tion mode, are provided in T able 6 and Figure 8 .
The overall accuracy of the LSTM model was 72.46%.
The model did well in the ’ train’ category , with a recall rate
of 91%, but struggled in the ’ taxi’ and ’ subway’ categories,
with recall rates of 24% and 37%, respectively . The details
of the LSTM performance, including precision, and recall
for each transportation mode, are provided in T able 7 and
Figure 9 .

124 Informatica 48 (2024) 1 17–130 S. Murrar et al.
T able 4: Precision, recall, and f1-score for MLP model.
Mode walk bike bus car taxi train subway accuracy pr ecision r ecall support
walk 268,077 22,360 8,847 467 1 2 0 89.43% 63% 89% 299,754
bike 30,229 144,562 12,078 2,350 34 66 40 76.34% 67% 76% 189,359
bus 78,596 25,694 136,840 2,776 603 8,129 1,895 53.76% 69% 54% 254,533
car 21,626 4,906 15,392 56,029 879 2,327 978 54.85% 81% 55% 102,137
taxi 10,702 8,664 13,1 1 1 2,488 3,809 8,683 1,076 7.84% 68% 8% 48,533
train 2,937 667 5,21 1 1,689 27 101,754 81 90.55% 81% 91% 1 12,366
subway 16,690 7,461 5,498 3,584 253 5,418 18,090 31.74% 82% 32% 56,994
T able 5: Precision, recall, and f1-score for KNN model.
Mode walk bike bus car taxi train subway accuracy pr ecision r ecall support
walk 259,771 8,869 23,096 3,076 1,497 385 3,060 86.66% 77% 87% 299,754
bike 14,327 160,364 1 1,1 16 1,719 787 139 907 84.68% 84% 85% 189,359
bus 40,956 15,132 183,604 3,602 3,917 4,798 2,524 72.13% 76% 72% 254,533
car 7,828 2,981 6,503 79,371 1,939 746 2,769 77.71% 82% 78% 102,137
taxi 5,431 1,866 8,707 3,392 24,025 4,194 918 49.50% 67% 50% 48,533
train 1,514 483 3,817 902 2,403 102,688 559 91.38% 89% 91% 1 12,366
subway 7,754 1,437 5,029 5,291 1,500 2,367 33,616 58.98% 76% 59% 56,994
The accuracy of the RNN model was 70.86%. It excelled
in the ’ train’ category , correctly identifying approximately
89% of instances. However , it performed poorly in the
’ taxi’ and ’ subway’ categories, with recall rates of 23% and
37%, respectively ( 8 and Figure 10 ).
The overall accuracy of the Logistic Regression model
was 50.99%. Most categories were dif ficult for the model
to distinguish, particularly ’ taxi,’ where it failed to identify
any instances correctly . Surprisingly , the model performed
relatively well in the ’walk’ and ’ train’ categories, correctly
identifying approximately 86% and 75% of instances, re-
spectively . The Logistic Regression model performance
details, including the precision and recall for each trans-
portation mode, are provided in T able 9 and Figure 1 1 .
According to previous results, the DT model emer ges as
Figure 9: Performance Comparison of T ransport Modes
with LSTM Algorithm.
the optimal choice in a comparative analysis of various ML
algorithms based on crucial evaluation metrics, as shown
in T able 10 , despite a slightly lower accuracy of 85%, as
opposed to the highest of 89% manifested by the MLP ,
LSTM, and RNN algorithms. The DT model is superior
because it has the lowest recorded bias of 16% and the low-
est competitive variance of 15%. These indicators point to
improved model robustness compared to its counterparts,
reducing the risk of overfitting or underfitting.
Importantly , in the context of T able 10 , the DT model
has 84% precision, indicating a lower probability of false-
positive instances. At the same time, it m aintains a com-
mendable recall rate of 85%, demonstrating its ef fective-
ness in identifying true positives. Furthermore, the DT al-
gorithm’ s F1-score, representing the harmonic mean of pre-
Figure 10: Performance Comparison of T ransport Modes
with RNN Algorithm.

Machine Learning Algorithms for T ransportation Mode… Informatica 48 (2024) 1 17–130 125
T able 6: Precision, recall, and f1-score for the DT model.
Mode walk bike bus car taxi train subway accuracy pr ecision r ecall support
walk 254,616 5,355 29,470 2,248 3,936 862 3,267 84.94% 84% 85% 299,754
bike 5,840 175,546 6,088 420 327 101 1,037 92.70% 93% 93% 189,359
bus 32,520 6,044 209,397 1,167 2,150 1,446 1,809 82.26% 82.26% 83% 254,533
car 2,81 1 457 1,208 94,927 569 279 1,886 92.94% 94% 93% 102,137
taxi 2,910 320 2,090 592 40,355 1,362 904 83.15% 81% 83% 48,533
train 785 78 1,515 235 1,407 107,869 477 96.00% 96% 96% 1 12,366
subway 4,987 384 1,926 1,088 1,058 456 47,095 82.63% 83% 83% 56,994
T able 7: Precision, recall, and f1-score for LSTM model.
Mode walk bike bus car taxi train subway accuracy pr ecision r ecall support
walk 266,177 15,745 13,846 3,686 196 25 79 88.38% 66% 89% 299,754
bike 24,629 147,263 14,329 2,804 155 72 107 78.91% 74% 78% 189,359
bus 69,971 17,356 152,815 2,622 3,002 6,413 2,354 59.47% 73% 60% 254,533
car 14,940 3,683 9,632 69,781 2,060 1,062 979 65.50% 78% 68% 102,137
taxi 7,902 8,769 9,620 3,023 1 1,472 7,033 714 22.60% 61% 24% 48,533
train 2,1 10 695 4,137 1,560 1,267 102,408 189 91.42% 85% 91% 1 12,366
subway 15,880 5,364 4,447 5,799 717 3,935 20,852 36.18% 83% 37% 56,994
cision and recall, peaks at 84%
Bike T ravel Mode Given the data in T able 1 1 , which
compares various ML algorithms for predicting bike travel
mode, it is clear that the DT model outperforms the others.
It achieves the highest accuracy of 93%, a significant ad-
vantage over the second best, the KNN algorithm, which
achieves 85%. Furthermore, the DT model is exceptionally
stable, with the lowest recorded bias and variance, which
are 7%. This implies that this model is less prone to overfit-
ting or underfitting, improving its overall reliability in bike
travel mode prediction.
The DT model outperforms all other models in preci-
sion, recall, and F1-score, critical measures in determining
a model’ s ef fectiveness at accurately predicting true posi-
tives and its balance of false positives and true positives.
Figure 1 1: Performance Comparison of T ransport Modes
with Logistic Regression Algorithm
Therefore, based on the comprehensive evaluation pre-
sented in T able 1 1 , the DT model appears to provide the
most beneficial trade-of f among accuracy , bias-variance
equilibrium, precision, recall, and F1-score in the context
of predicting bike travel mode.
Bus T ravel Mode As presented in T able 12 , the DT
model outperforms the other ML algorithms in bus travel
mode prediction. W ith an accuracy rate of 82%, it signif-
icantly surpasses the second-best performer , KNN, which
achieves 72% accuracy . The DT model demonstrates re-
markable robustness, evident in its lowest recorded bias of
17% and equally commendable variance rate of 18%. Addi-
tionally , the model excels in precision, recall, and F1-score,
measuring at 83%. These results underscore the model’ s
superior ability to predict true positives and ef fectively bal-
ance false positives and true positives.
Therefore, based on the comprehensive evaluation pre-
sented in T able 12 , the DT model emer ges as the optimal
choice for bus travel mode prediction, of fering the best
trade-of f between accuracy , bias-variance balance, preci-
sion, recall, and F1-score.
Car T ravel Mode Based on the data presented in T able
13 for car travel mode prediction, the DT model demon-
strates exceptional performance again compared to the
other evaluated ML models. W ith an accuracy rate of 93%,
it significantly outperforms the second-best model, KNN,
which achieves an accuracy rate of 78%. The robustness
of the DT model is further highlighted by its minimal bias
of 6% and remarkably low variance of 7%, indicating a re-
duced likelihood of overfitting or underfitting and enhanc-
ing the algorithm’ s overall reliability .
Furthermore, the DT model excels in precision, recall,
and F1-score, achieving a score of 94% in each category .
This reflects its ability to accurately predict true positives
while maintaining a balanced proportion of false positives.
In conclusion, the comprehensive evaluation presented in

126 Informatica 48 (2024) 1 17–130 S. Murrar et al.
T able 8: Precision, recall, and f1-score for RNN model.
Mode walk bike bus car taxi train subway accuracy pr ecision r ecall support
walk 260,588 16,829 17,572 4,395 108 79 183 89.40% 66% 87% 299,754
bike 26,734 143,799 14,995 3,059 382 46 344 73.72% 72% 76% 189,359
bus 67,050 20,776 149,221 4,603 3,873 5,868 3,142 57.70% 71% 59% 254,533
car 14,682 4,133 10,155 67,852 2,806 722 1,787 63.52% 75% 66% 102,137
taxi 8,712 9,012 9,222 2,908 1 1,126 6,565 988 19.56% 53% 23% 48,533
train 2,227 722 4,992 1,886 1,790 99,845 904 89.99% 85% 89% 1 12,366
subway 15,791 5,225 4,380 5,554 752 3,971 21,321 35.65% 74% 37% 56,994
T able 9: precision, recall, and f1-score for the Logistic Regression model.
Mode walk bike bus car taxi train subway accuracy pr ecision r ecall support
walk 258,048 21,357 20,290 54 0 5 0 86.08% 54% 86% 299,754
bike 69,838 49,492 66,581 3,425 0 23 0 26.13% 51% 26% 189,359
bus 84,460 20,941 126,863 1 1,361 1 8,738 2,169 49.84% 42% 50% 254,533
car 26,056 1,732 42,307 17,882 0 13,679 481 17.50% 30% 18% 102,137
taxi 15,71 1 1,152 17,547 10,610 0 3,422 91 0% 0% 0% 48,533
train 2,292 256 15,71 1 9,867 0 84,233 7 74.96% 73% 75% 1 12,366
subway 21,909 1,626 16,1 15 6,050 0 5,345 5,949 10.43% 68% 10% 56,994
T able 10: Performance of V arious ML Algorithms for W alk Mode Prediction.
Algorithm Accuracy Bias V ariance Pr ecision Recall F1-scor e
MLP 89% 38% 9% 63% 89% 74%
KNN 87% 23% 13% 77% 87% 82%
DT 85% 16% 15% 84% 85% 84%
LSTM 88% 34% 12% 66% 89% 76%
RNN 89% 35% 1 1% 66% 87% 75%
Logistic Regression 86% 46% 14% 54% 86% 66%
T able 13 solidifies the DT model as the optimal choice for
car travel mode prediction, providing the most favorable
trade-of f among accuracy , bias-variance balance, precision,
recall, and F1-score.
Subway T ravel Mode As shown in T able 14 , the DT
model outperforms the other ML algorithms under con-
sideration for predicting subway travel mode. It has an
astounding accuracy rate of 83%, far outperforming the
second-best-performing algorithm, KNN, which has an ac-
curacy rate of 59%. The DT algorithm’ s bias and variance
scores of 17% further demonstrate its robustness. These re-
sults indicate that the model has an impressive robustness
that reduces the likelihood of overfitting or underfitting,
thereby increasing its reliability for this prediction task.
Furthermore, the DT model outperforms precision, re-
call, and F1-score, scoring 83%
T axi T ravel Mode According to the analysis of the data
in T able 15 for taxi travel mode prediction, the DT model
outperforms all other evaluated ML models significantly .
The DT model has an accuracy rate of 83%, which is con-
siderably higher than the next most accurate model, KNN,
which has a rate of 50%. The DT model’ s bias and variance
rates, both less than 20%, highlight its exceptional robust-
ness, implying a lower propensity for overfitting or under -
fitting, thus contributing to overall model reliability .
Furthermore, the DT model performs admirably in pre-
cision, recall, and F1-score, with scores of 81%
T rain T ravel Mode According to T able 16 , which com-
pares ML models for predicting train travel mode, the DT
model is superior . The DT model has the highest accuracy
of 96%, outperforming the MLP , KNN, LSTM, and RNN
models, all of which have accuracies in the lower nineties.
The DT model also demonstrates superior robustness, with
a recorded bias of 4% and a variance rate of 4%, imply-
ing less susceptibility to overfitting or underfitting and thus
increasing its reliability for the prediction task.
The DT model outperforms its competitors in precision,
recall, and F1-score, scoring 96% across all three metrics.
These scores represent not only the model’ s ability to iden-
tify true positives accurately but also its ef fectiveness in
maintaining a balance between true positives and false pos-
itives. As a result of the comprehensive evaluation in T able
16 , the DT model can be considered the best choice for pre-
dicting train travel mode.
5.1 Comparative analysis of computational
complexity
After conducting a comprehensive analysis of key perfor -
mance metrics, such as accuracy , precision, recall, and F1-
score, for various transportation modes (including walking,
biking, taking the bus, driving a car , taking a taxi, riding

Machine Learning Algorithms for T ransportation Mode… Informatica 48 (2024) 1 17–130 127
T able 1 1: Performance of V arious ML Algorithms for Bike Mode Prediction.
Algorithm Accuracy Bias V ariance Pr ecision Recall F1-scor e
MLP 76% 30% 25% 67% 76% 72%
KNN 85% 16% 15% 84% 85% 84%
DT 93% 7% 7% 93% 93% 93%
LSTM 79% 27% 21% 74% 78% 76%
RNN 74% 27% 26% 72% 76% 74%
Logistic Regression 26% 49% 74% 51% 26% 35%
T able 12: Performance of V arious ML Algorithms for Bus Mode Prediction.
Algorithm Accuracy Bias V ariance Pr ecision Recall F1-scor e
MLP 54% 33% 45% 69% 54% 61%
KNN 72% 24% 28% 76% 72% 74%
DT 82% 17% 18% 83% 82% 83%
LSTM 59% 26% 41% 73% 60% 66%
RNN 58% 30% 42% 71% 59% 64%
Logistic Regression 50% 58% 50% 42% 50% 45%
T able 13: Performance of V arious ML Algorithms for Car Mode Prediction.
Algorithm Accuracy Bias V ariance Pr ecision Recall F1-scor e
MLP 55% 17% 55% 81% 55% 65%
KNN 78% 18% 22% 82% 78% 80%
DT 93% 6% 7% 94% 93% 94%
LSTM 66% 21% 34% 78% 68% 73%
RNN 64% 25% 36% 75% 66% 71%
Logistic Regression 18% 70% 82% 30% 18% 22%
T able 14: Performance of V arious ML Algorithms for Subway Mode Prediction.
Algorithm Accuracy Bias V ariance Pr ecision Recall F1-scor e
MLP 32% 18% 69% 82% 32% 46%
KNN 59% 24% 41% 76% 59% 66%
DT 83% 17% 17% 83% 83% 83%
LSTM 36% 18% 64% 83% 37% 51%
RNN 36% 23% 64% 74% 37% 50%
Logistic Regression 10% 32% 90% 68% 10% 18%
T able 15: Performance of V arious ML Algorithms for T axi Mode Prediction.
Algorithm Accuracy Bias V ariance Pr ecision Recall F1-scor e
MLP 8% 37% 90% 68% 8% 14%
KNN 50% 33% 50% 67% 50% 57%
DT 83% 19% 17% 81% 83% 82%
LSTM 23% 39% 77% 61% 24% 34%
RNN 20% 41% 80% 53% 23% 32%
Logistic Regression 0% 100% 100% 0% 0% 0%
the train, and using the subway), it is essential to choose
an algorithm that aligns with the specific requirements of
the practical application. This alignment entails balancing
the availability of computational resources with the require-
ment for accuracy and complexity in predictions.
It is equally crucial to consider each algorithm’ s com-
putational complexity . This factor is vital, particularly in
practical situations where there are limitations on computa-
tional resources. T able 17 presents a comparative analysis
of the computational complexity for each model.
This analysis emphasizes various crucial factors to con-
sider when choosing a suitable ML algorithm:

128 Informatica 48 (2024) 1 17–130 S. Murrar et al.
T able 16: Performance of V arious ML Algorithms for T rain Mode Prediction.
Algorithm Accuracy Bias V ariance Pr ecision Recall F1-scor e
MLP 91% 21% 9% 81% 91% 85%
KNN 91% 1 1% 9% 89% 91% 90%
DT 96% 4% 4% 96% 96% 96%
LSTM 91% 17% 9% 85% 91% 88%
RNN 90% 16% 10% 85% 89% 87%
Logistic Regression 75% 27% 25% 73% 75% 74%
T able 17: Comparative Analysis of Computational Complexity of V arious ML Algorithms.
Algorithm Complexity T raining T ime Resour ce Intensity
MLP High (multiple layers) Longer (complex) High (resource-intensive)
KNN Low to Moderate (lazy) Minimal (higher in predic-
tion)
High (lar ge datasets)
DT Moderate (depends on
depth)
Faster (simpler) Lower (ef ficient)
LSTM High (complex RNN vari-
ant)
Long (detailed architecture) High (resource-heavy)
RNN High (sequential loops) Lengthy (for sequences) High (intensive for longer
sequences)
Logistic Regression Low (linear model) Shorter (less parameters) Low (ef ficient for simpler
tasks)
– For Limited Computational Resources: Logistic Re-
gression and Decision T rees are preferable due to their
lower complexity and resource requirements. These
models are ideal for applications with constrained
computational capacity .
– For Higher Accuracy and Complex Patterns: MLP ,
LSTM, and RNN are better suited, albeit at the cost
of higher computational resources and longer training
times. These models are advantageous in scenarios
where accuracy is critical and complex patterns are
present in the data.
– Balance between Accuracy and Computational Ef fi-
ciency: KNN might be a good middle ground. How-
ever , it is essential to note that KNN can be less ef fi-
cient for lar ge datasets due to its high resource inten-
sity during the prediction phase.
6 Discussion
The present study makes a noteworthy contribution to the
transportation mode prediction field by utilizing ML algo-
rithms. This research examines the ef fectiveness of dif fer -
ent algorithms, with a specific focus on the DT model. It
provides fresh perspectives and raises questions about cur -
rent practices in using ML for transportation analytics.
The study’ s results emphasize the exceptional accuracy
of the DT model in forecasting transportation modes based
on extensive datasets. It achieved precision, recall, and
F1 scores between 83% and 94% for all transportation cat-
egories. This performance stands out compared to simi-
lar works, which primarily employed more intricate mod-
els such as Convolutional Neural Networks (CNNs), Long-
term Recurrent Convolutional Networks (LRCNs), and
other advanced deep learning techniques.
Previous studies used dif ferent approaches to tackle the
intricacies of predicting travel behavior . For example,
Convolutional Neural Networks (CNNs), renowned for
their ability to extract features, have demonstrated ef fec-
tiveness but necessitated a lar ge amount of data and de-
manded significant computational resources. Deep Neu-
ral Networks (DNNs) have shown their ef fectiveness in
learning long sequences, but they are susceptible to over -
fitting and require substantial computational resources.
Long- and short-term recurrent convolutional networks
(LRCNs), which mer ge the advantages of CNNs and RNNs,
have been determined to be well-suited for analyzing se-
quential spatial data. Alternative methodologies such as
the Spatial-T emporal Pattern Chain Network (STPC-Net)
and Sequence-to-sequence models have also demonstrated
notable precision. However , their intricate nature has
prompted concerns regarding their feasibility for less com-
plex tasks.
The current study is notable for its ability to showcase
the ef ficacy of the DT model. This discovery is especially
significant considering the model’ s comparatively straight-
forward nature compared to the more intricate models typ-
ically employed in similar studies. The exceptional perfor -
mance of the DT model contradicts the current inclination
towards more intricate solutions in the field of transporta-
tion mode prediction. It indicates that less complex models,
which are more ef ficient in computation and easier to un-
derstand, can ef fectively handle the intricacies of predicting

Machine Learning Algorithms for T ransportation Mode… Informatica 48 (2024) 1 17–130 129
travel behavior .
Examining why specific models exhibited superior per -
formance in this study is centered on various factors. The
DT model’ s capacity to attain high precision without requir -
ing abundant data and computational resources represents
a notable advantage, mainly when it is scarce. The model’ s
high precision, recall, and F1 scores demonstrate its robust-
ness across dif ferent transportation modes. This character -
istic may not be as prominent in more intricate models re-
quiring meticulous adjustments and extensive training data.
The findings have significant implications for the field of
transportation analytics. They propose a possible transition
towards more ef fective and adaptable solutions, empha-
sizing the significance of weighing the trade-of f between
model intricacy and performance ef fectiveness. This ap-
proach has the potential to create analytical tools for pre-
dicting transportation modes that are both accessible and
sustainable. This would be advantageous for researchers
and practitioners who have limited resources.
This study enhances the existing literature by emphasiz-
ing the capability of simpler ML models, such as Decision
T rees, to forecast transportation modes more ef ficiently and
accurately . It provides opportunities for future research to
investigate models that balance balance and practical ap-
plicability . This could potentially result in more accessible
and sustainable solutions in the field.
7 Conclusion
In conclusion, our research has made significant strides
in exploring the application of various machine learning
(ML) techniques, such as Multilayer Perceptron (MLP),
K-Nearest Neighbors (KNN), Decision T ree, Long Short-
T erm Memory (LSTM), Recurrent Neural Network (RNN),
and Logistic Regression, for accurately predicting trans-
portation modes like cars, bikes, and buses. The Decision
T ree (DT) method has demonstrated notable ef fectiveness
due to its accuracy , simplicity , and adaptability . These find-
ings are particularly relevant for enhancing urban planning
and traf fic management, promising to improve traf fic flow
and the ef ficiency of public transportation systems. While
methods like MLP and LSTM have their limitations, they
still hold value for applications in travel apps, of fering per -
sonalized route suggestions.
However , our study acknowledges several limitations,
including computational and scalability challenges with
complex models, the influence of temporal and seasonal
factors on transportation patterns, data privacy and secu-
rity concerns, sensor accuracy , and the cultural and regional
applicability of the models. These constraints highlight the
need for further research in this field. T o this end, it is es-
sential to address these limitations to harness the potential
of ML fully in transportation mode prediction. Future re-
search should incorporate more integration models to refine
the accuracy and reliability of predictions across all trans-
portation modes. This approach will advance the ML field
in transportation and contribute significantly to developing
more intelligent, more ef ficient urban environments.
Acknowledgment
This research work was made possible through a fund from
Applied Science Private University in Amman, Jordan, to
cover part of the publication fee.
Refer ences
[1] J. Zeng, Y . Y u, Y . Chen, D. Y ang, L. Zhang,
D. W ang, T rajectory-as-a-Sequence: A novel travel
mode identification framework, T ransportation Re-
sear ch Part C: Emer ging T echnologies , 146, pp.
103957, 2023. DOI: https://doi.org/10.1016/
j.trc.2022.103957
[2] O. Bagdadi, A. Várhelyi, Development of a method
for detecting jerks in safety critical events, Ac-
cident Analysis & Pr evention , 50, pp. 83-91,
2013. DOI: https://doi.org/10.1016/j.aap.
2012.03.032
[3] H. Mäenpää, A. Lobov , J. Lastra, T ravel mode esti-
mation for multi-modal journey planner , T ransporta-
tion Resear ch Part C: Emer ging T echnologies , 82, pp.
273-289, 2017. DOI: https://doi.org/10.1016/
j.trc.2017.06.021
[4] C. Hasif, A. Araldo, S. Dumbrava, D. W atel, A graph-
database approach to assess the impact of demand-
responsive services on public transit accessibility ,
In pr oc. of 15th ACM SIGSP A TIAL International
W orkshop on Computational T ransportation Science,
pp. 1-4, 2022. DOI: https://doi.org/10.1145/
3557991.3567798
[5] C. Hof fmann, C. Abraham, M. White, S.
Ball, S. Skippon, What cognitive mecha-
nisms predict travel mode choice? A system-
atic review with meta-analysis, T ransport Re-
views , 37(5), pp. 631-652, 2017. DOI: https:
//doi.org/10.1080/01441647.2017.1285819
[6] T . Phiophuead, N. Kunsuwan, Logistic regression
analysis of factors af fecting travel mode choice for
disaster evacuation, Engineering Journal , 23(6), pp.
399-417, 2019. DOI: https://doi.org/10.4186/
ej.2019.23.6.399
[7] Y . Zheng, L. Zhang, X. Xie, W . Ma, Mining inter -
esting locations and travel sequences from GPS tra-
jectories, In Pr oc. of 18th international conference on
W orld wide web, pp. 791-800, 2009. DOI: https:
//doi.org/10.1145/1526709.1526816

130 Informatica 48 (2024) 1 17–130 S. Murrar et al.
[8] Y . Zheng, Q. Li, Y . Chen, X. Xie, W . Ma, Under -
standing mobility based on GPS data . In Pr oc. of
10th international conference on Ubiquitous comput-
ing, pp. 312-321, 2008. DOI: https://doi.org/
10.1145/1409635.1409677
[9] S. Ding, Z. Li, K. Zhang, F . Mao, A Comparative
Study of Frequent Pattern Mining with T rajectory
Data, Sensors , 22 (19), pp. 7608, 2022. DOI: https:
//doi.org/10.3390/s22197608 .
[10] S. Dabiri, K. Heaslip, Inferring transportation modes
from GPS trajectories using a convolutional neural
network, T ransportation Resear ch Part C: Emer ging
T echnologies , 86, pp. 360-371, 2018. DOI: https:
//doi.org/10.1016/j.trc.2017.11.021 .
[1 1] J. Y u, T ravel Mode Identification W ith GPS T rajec-
tories Using W avelet T ransform and Deep Learning,
IEEE T ransactions on Intelligent T ransportation Sys-
tems , 22(2), pp. 1093-1 103, 2021. DOI: https://
doi.org/10.1109/TITS.2019.2962741
[12] Y . Zhu, Y . Liu, J. Y u, X. Y uan, Semi-Supervised
Federated Learning for T ravel Mode Identification
From GPS T rajectories, IEEE T ransactions on In-
telligent T ransportation Systems , 23(3), pp. 2380-
2391, 2022. DOI: https://doi.org/10.1109/
TITS.2021.3092015
[13] J. Kim, J. Kim, G. Lee, GPS data-based mo-
bility mode inference model using long-term re-
current convolutional networks, T ransportation Re-
sear ch Part C: Emer ging T echnologies , 135, pp.
103523, 2022. DOI: https://doi.org/10.1016/
j.trc.2021.103523
[14] C. Zheng, C. W ang, X. Fan, J. Qi, X. Y an, STPC-
Net: Learn Massive Geo-Sensory Data as Spatio-
T emporal Point Clouds, IEEE T ransactions on In-
telligent T ransportation Systems , 23(8), pp. 1 1314-
1 1324, 2022. DOI: https://doi.org/10.1109/
TITS.2021.3102747
[15] W . Noble, What is a support vector machine?, Natur e
biotechnology , 24(12), pp. 1565-1567, 2006. DOI:
https://doi.org/10.1038/nbt1206- 1565
[16] J. Brownlee, Long short-term memory networks with
python: develop sequence prediction models with
deep learning, Machine Learning Mastery , 2017.
[17] R. Pascanu, C. Gulcehre, K. Cho, Y . Bengio, How
to construct deep recurrent neural networks, arXiv
pr eprint arXiv:1312.6026 , 2013. DOI: https://
doi.org/10.48550/arXiv.1312.6026
[18] Y . Izza, A. Ignatiev , J. Marques-Silva, On explain-
ing decision trees. arXiv pr eprint arXiv:2010.1 1034 ,
2020. DOI: https://doi.org/10.48550/arXiv.
2010.11034
[19] M. Machado, S. Karray , I. de Sousa, LightGBM:
An ef fective decision tree gradient boosting method
to predict customer loyalty in the finance industry ,
In Pr oc. of 14th International Conference on Com-
puter Science & Education (ICCSE), pp. 1 1 1 1-1 1 16,
T oronto, ON, Canada, 2019. DOI: https://doi.
org/10.1109/ICCSE.2019.8845529
[20] L. Noriega, Multilayer perceptron tutorial, School
of Computing, Staffor dshir e University , 4, pp.
5, 2005. DOI: https://api.semanticscholar.
org/CorpusID:61645526
[21] X. Zou, Y . Hu, Z. T ian, K. Shen, Logistic regres-
sion model optimization and case analysis, In Pr oc.
of IEEE 7th International Conference on Computer
Science and Network T echnology (ICCSNT), pp.
135-139, 2019. DOI: https://doi.org/10.1109/
ICCSNT47585.2019.8962457
[22] D. Larose, C. Larose, k-Nearest Neighbor Algorithm,
Discovering Knowledge in Data: An Intr oduction to
Data Mining , pp. 149-164, 2014. DOI: https://
doi.org/10.1002/9781118874059.ch7
[23] M. Shulajkovska, G. Noveski, M. Smerkol, J. Grab-
nar , E. Dovgan, M. Gams, EU Smart Cities: T owards
a New Framework of Urban Digital T ransformation,
Informatica , (2), pp. 151-158, 2023. DOI: https:
//doi.org/10.31449/inf.v47i2.4904
[24] Z. He, Improved Genetic Algorithm in Multi-
objective Car go Logistics Loading and Distribution,
Informatica , 47(2), pp. 253-260, 2023. DOI: https:
//doi.org/10.31449/inf.v47i2.3958