Content uploaded by Leo Tisljaric
Author content
All content in this area was uploaded by Leo Tisljaric on Oct 05, 2021
Content may be subject to copyright.
Classification of Travel Modes from Cellular
Network Data Using Machine Learning Algorithms
Leo Tišljari´
c1, Dominik Cvetek1, Valentin Vareški´
c2, Martin Greguri´
c1
1Faculty of Transport and Traffic Sciences, University of Zagreb
Vukeli´
ceva 4, HR-10000 Zagreb, Croatia
2iOLAP d.o.o.
Prolaz Marije Krucifikse Kozuli´
c 1, HR-51000, Rijeka, Croatia
ltisljaric@fpz.unizg.hr
Abstract—Data availability in recent years has grown expo-
nentially, allowing researchers in the transport sector to harness
valuable information regarding traffic flows. In that sense, cel-
lular network data represents valuable traffic information when
dealing with spatially large areas due to its property of collecting
route data using distant mobile base stations. This property
enables the automatic collection of origin-destination data, which
is traditionally collected using field or online questionnaires. This
paper aims to present the possibility of using origin-destination
data extracted from cellular network dataset to classify travel
modes. A case study was performed on the dataset collected in
the City of Rijeka, Croatia. Dataset is evaluated on five machine
learning algorithms, which resulted in Random forest as the
highest performing algorithm with an accuracy score of 99.93%.
Keywords—Travel mode classification; Machine learning; Cel-
lular network data; Origin-destination matrices
I. INTROD UC TI ON
Deployment of various traffic sensors combined with the
increased development of data processing techniques and com-
puting power resulted in exponential growth of available traffic
datasets in the recent decade. Consequently, many data-driven
road traffic-related research topics emerged like traffic data
fusion models [1], [2], traffic state estimation and prediction
[3], [4], and traffic control [5].
In this paper, the methodology for the classification of
Origin-Destination (OD) data extracted from cellular network
data using machine learning algorithms is presented. The
paper’s primary goal is to compare the performances of the
most used machine learning algorithms to identify and propose
the best-performing ones.
Authors in [6] compared classification algorithms for se-
lecting the travel mode. Random Forest (RF) and K-Nearest
Neighbors (KNN) were compared. RF showed better accuracy
for travel mode classification applied to Global Navigation
Satellite System (GNSS) historical data and data streaming.
Zang et al. in [7] analyzed the navigational behavior of users
using GNSS traces. The goal of analyzing data is to enable
the construction of a more fine-tuned optimal route. Authors in
[8] discuss a new travel mode classification approach using a
Convolutional Neural Network (CNN) on accelerometer data.
Authors in [9] analyze the potential and limitations of using
cellular network data for traffic analysis. Katatian et al. [10]
calculate travel times as the primary attributes for clustering.
Trips between the exact origin and destination zones are
combined in a cluster. KNN is used to calculate clusters
representing a particular travel modes: walking, public trans-
portation, or driving a private car. Authors in [11] compare
three geometry-based mode classification methods and three
supervised methods to classify trips extracted from the cellular
network.
Contributions of this paper are as follows: (i) methodol-
ogy for processing OD-based datasets for machine learning
algorithms, (ii) evaluation of machine learning algorithms for
classification of travel modes using OD dataset, and (iii)
proposed methodology and evaluation are applied on the real-
life dataset from the City of Rijeka, Croatia.
The rest of the paper is organized as follows. Section 2
presents a used methodology that describes the usage of OD
matrices, presents the used dataset, prepossessing steps, and
feature selection part. Then, every used machine learning
algorithm is described with advantages and disadvantages.
Section 3 presents the results of the evaluation on five machine
learning algorithms. The conclusion and future work directions
are given in section 4.
II. METHODOLOGY
A. Origin-destination matrix
Cells in the OD matrix represent the number of trip records
in the observed time period, where each trip is realized as
unimodal. Fig.1 represents the used OD matrix where the
trips are recorded at the City of Rijeka, Croatia. Data was
collected and aggregated for average working day in a year
with excluded data from a holiday season. The data contains
records divided into 48 sectors covering the wide city area
represented in Fig.2.
The OD matrix is mostly used as an input for the traffic
simulations, where the matrix is used for generating the trips
that represent the initial traffic model of the observed area
[12]. The matrix can also be used as an input for traffic state
estimation, anomaly detection or prediction models because it
highlights the area of the increased traffic activity [13].
978-1-6654-4437-8/21/$31.00 ©2021 IEEE
63rd International Symposium ELMAR-2021, 13-15 September 2021, Zadar, Croatia
173
2021 International Symposium ELMAR | 978-1-6654-4437-8/21/$31.00 ©2021 IEEE | DOI: 10.1109/ELMAR52657.2021.9550817
Authorized licensed use limited to: University of Zagreb: Faculty of Electrical Engineering and Computing. Downloaded on October 05,2021 at 11:35:13 UTC from IEEE Xplore. Restrictions apply.
Attribute Description
Interval Time interval when the trip started.
Duration Duration of the trip [s].
Start sector Id of the trip starting sector.
End sector Id of the trip ending sector.
Air distance Euclidean distance between start
and end of the trip [m].
Air speed Obtained speed using Euclidean
distance [m/s].
Road distance Real road distance between start
and end of the trip [m].
Road speed Obtained speed using road distance
[m/s].
Mode Travel mode used for the trip.
TABLE I. Attributes of the dataset
B. Data
The used dataset contains around 500,000 OD records
extracted from the cellular network, which represent trip data.
Every trip record is described with nine attributes presented
in Table I. Attribute ’Interval’ represents the time interval in
which the trip is recorded, with possible values of 00:00-06:00,
06:00-09:00, 09:00-14:00, 14:00-18:00 or 18:00-00:00h. ’Du-
ration’ represents the total trip duration in seconds. The
’Air distance’ and ’Air speed’ represent computed Euclidean
distance and the speed calculated using Euclidean distance
and the difference in time from the trip start to end. The
’Road distance’ and ’Road speed’ represent computed actual
distance and speed of vehicles when traveling using road
infrastructure. The ’Mode’ attribute represents the travel mode
used for completing the trip, where recorded travel modes are
’car,’ ’public transport,’ and ’walking.’
C. Data preprocessing
Distributions of the observed attributes are shown in the
diagonal images in Fig.3. Due to highly skewed distributions,
we used the adjusted box plot method to remove anomalies
because it is a method that does not take any parametric as-
sumptions and uses med couple as a robust skewness estimator
[14]. The anomaly detection resulted in excluding trips that
had a duration longer than 9,000 s (2.5h).
Figure 1. OD matrix for recorded trips at the City of Rijeka and its nearby
surrounding
Figure 2. Sectors (spatial zones) for the City of Rijeka
After the anomaly detection, the feature scaling step is
conducted. To mitigate the influence of different units and
large differences in attribute values, all attributes are scaled
to the [0,1] range.
After the data preprocessing, 504,419 trip records were used
for further analysis. The most dominant travel mode was a
’car’ with 398,729 records, followed by ’public transport’ with
83,060 records, and ’walking’ with 22,630 records.
D. Feature selection
After examining all dataset attributes represented in Table I,
four features were selected for further analysis and the learning
process: ’Duration,’ ’Air speed,’ ’Road distance,’ and ’Road
speed.’
When considering the correlation plots in Fig.3, it can be
observed that attributes ’Road distance’ and ’Air distance’
provide redundant information, and one can be removed. We
removed the ’Air distance’ attribute from the learning because
road distance is a more informative attribute for traffic data
analysis.
E. Machine learning algorithms
This section presents five machine learning algorithms
for classification of the travel modes used in this research
with corresponding advantages and disadvantages. We analyze
following algorithms: (i) Decision tree (DT), (ii) K-Nearest
Neighbors, (iii) Logistics Regression (LR), (iv) Naive Bayes
(NB), and (v) Random Forest.
1) K-Nearest Neighbors: Machine learning algorithm that
solves classification and regression problems. KNN is a non-
parametric classification method, which classifies data points
based on its similarity measure. Algorithm calculates the dis-
tance between data points using the preferred distance metric
and adds the distance to an ordered collection. Then, it sort the
ordered collection from smallest to largest and picks the first K
groups from the sorted collection [15]. The main disadvantage
of KNN as it becomes significantly slower with increasing data
volume makes it an impractical choice in environments where
rapid forecasting is required. The main advantages of KNN for
63rd International Symposium ELMAR-2021, 13-15 September 2021, Zadar, Croatia
174
Authorized licensed use limited to: University of Zagreb: Faculty of Electrical Engineering and Computing. Downloaded on October 05,2021 at 11:35:13 UTC from IEEE Xplore. Restrictions apply.
classification are very simple application, robust in terms of
search space; for example, classes do not have to be linearly
separable.
2) Decision Tree: Machine learning algorithm that solves
both classification and regression tasks. A DT can be used to
visually represent the decision-making process using a tree-
like model of decisions. The main advantages of a DT are
that it does not require scaling and normalization, and it is
very intuitive and easy to explain. The main disadvantage is
that it often involves a higher time to train the model, and
small changes in the data can lead to a large change in the
structure of the optimal decision tree because calculation can
go far more complex compared to other algorithms [16].
3) Random Forest: Machine learning algorithm that can
solve both classification and regression problems. RF builds
multiple decision trees and merges them together to get a
more accurate and stable predictions [17]. One of the benefits
of using the RF is the power of handling large data sets
with higher dimensionality, and it is an effective method
for estimating missing data. The main limitation is that a
large number of trees can make the algorithm too slow and
ineffective for real-time predictions.
4) Logistic regression: In a classification problem, the
output variable can only take discrete values for a given set of
inputs, and it models the data using the sigmoid function. The
advantage is that LR is easy to implement, interpret, and very
efficient to train. The disadvantage is that it is challenging to
capture complex relationships. In high-dimensional datasets,
this can lead to the model over-fitting into the training set.
Non-linear problems cannot be solved with LR because it has
a linear decision background [18].
5) Naive Bayes: NB is a classification technique based
on Bayes’ Theorem with assumption of independence among
predictors. It is used to discriminate against different objects
based on specific features. The main advantage of NB is that it
requires a small amount of training data. When the assumption
Figure 3. Distributions and correlations between attributes
63rd International Symposium ELMAR-2021, 13-15 September 2021, Zadar, Croatia
175
Authorized licensed use limited to: University of Zagreb: Faculty of Electrical Engineering and Computing. Downloaded on October 05,2021 at 11:35:13 UTC from IEEE Xplore. Restrictions apply.
Figure 4. Total accuracy for machine learning algorithms
of independent predictors holds true, the classifier performs
better as compared to other models. The main limitation
of NB is the assumption of independent predictors. In real
life scenario, the predictors are dependent, and it is almost
impossible that we get a set of predictors which are entirely
independent [19].
III. RESULTS
This section presents the evaluation of the cellular network
dataset on five machine learning algorithms. For each algo-
rithm, total accuracy is presented, alongside precision, recall,
F-1 score, and corresponding confusion matrices. The input
dataset is divided into training and test sets by using the
standard ratio of 30% for test and 70% for training. The
experiments are done using Python programming language
with the package Scikit-Learn [20]. The used code for this
research is publicly available on the Github repository [21].
The total accuracy of the algorithms is shown in Fig.4. It
can be observed that RF achieved the best result with a total
accuracy of 99.35%. KNN and DT achieved high accuracy
scores, while LR and NB failed to achieve sufficient scores.
Figure 5. Precision, recall and F1 score values for machine learning algorithms
Fig.5 represent precision, recall, and F-1 values for every
algorithm. We report the precision calculated as T P /(T P +
F P ), recall T P /(T P +F N ), and F-1 as a harmonic mean
between precision and recall, where T P stands for true
positive, F P for true negative, and F N for false negative
values.
Generally, confusion matrices, also called error matrices,
show the difference between a number of true versus predicted
class labels. Fig.6 represents confusion matrices with normal-
ized values for all observed algorithms. Confusion matrices
also confirm RF as the best performing algorithm with well-
separated classes. KNN and DT algorithms achieved good
separation between classes, NB could not separate the ’car’
and ’public transport’ classes, while LR could not separate
the ’car’ from ’public transport’ and ’walking’ from ’public
transport’ classes.
IV. CON CL US IO N
This paper presents a methodology for processing the OD
dataset extracted from the cellular network mobile records.
The paper’s main goal was to evaluate the dataset on the
most used machine learning algorithms for classification tasks
and report the results by comparing the performances. Based
on the results, conclusions can be drawn: (i) best performing
algorithm was RF, with a total accuracy of 99.93%, (ii) KNN
and DT algorithms can be used for this purpose because of
high accuracy rates and well separation of the classes, and (iii)
LR and NB are not well suited for the classification task on
this dataset.
Future work directions based on this dataset include auto-
matic feature extraction from OD matrices. As OD matrix is
presented as a heatmap, it can be used as an input for learning
the CNN to estimate the traffic states, and the results of the
CNN can be validated using the methodology proposed in this
paper.
ACKNOW LE DG ME NT
This research has been supported by the University of
Zagreb, Student Centre as part of the project “Znanstveno-
istraživaˇ
cke aktivnosti studentske istraživaˇ
cke skupine SIS-
DVA” and European Regional Development Fund under the
grant KK.01.1.1.01.0009 (DATACROSS). Data used for this
research is provided by Ericsson Nikola Tesla Ltd. through the
collaboration with the Laboratory for Data Science in Traffic
and Logistics at Faculty of Transport and Traffic Sciences,
University of Zagreb.
REFERENCES
[1] D. Cvetek, M. Muštra, N. Jeluši´
c, and L. Tišljari´
c, “A survey of methods
and technologies for congestion estimation based on multisource data
fusion,” Applied Sciences, vol. 11, no. 5, 2021.
[2] D. Cvetek, I. Horenec, M. Muštra, and N. Jeluši´
c, “Analysis of cor-
relation between dwell time measured using bluetooth detector and
occupancy,” in 2019 International Symposium ELMAR, pp. 31–34, 2019.
[3] L. Tišljari´
c, T. Cari´
c, B. Abramovi´
c, and T. Fratrovi´
c, “Traffic state
estimation and classification on citywide scale using speed transition
matrices,” Sustainability, vol. 12, no. 18, 2020.
63rd International Symposium ELMAR-2021, 13-15 September 2021, Zadar, Croatia
176
Authorized licensed use limited to: University of Zagreb: Faculty of Electrical Engineering and Computing. Downloaded on October 05,2021 at 11:35:13 UTC from IEEE Xplore. Restrictions apply.
(a) (b)
(c) (d)
(e)
Figure 6. Confusion matrices for machine learning algorithms: (a) Decision
tree, (b) K-nearest neighbors, (c) Logistics regression, (d) Naive Bayes, and
(e) Random forest
[4] T. Erdeli´
c, T. Cari´
c, M. Erdeli´
c, L. Tišljari´
c, A. Turkovi´
c, and N. Jeluši´
c,
“Estimating congestion zones and travel time indexes based on the
floating car data,” Computers, Environment and Urban Systems, vol. 87,
p. 101604, 2021.
[5] F. Vrbani´
c, E. Ivanjko, K. Kuši´
c, and D. ˇ
Cakija, “Variable speed limit
and ramp metering for mixed traffic flows: A review and open questions,”
Applied Sciences, vol. 11, no. 6, 2021.
[6] M. Erdeli´
c, T. Cari´
c, E. Ivanjko, and N. Jeluši´
c, “Classification of travel
modes using streaming gnss data,” Transportation Research Procedia,
vol. 40, pp. 209–216, 2019. TRANSCOM 2019 13th International
Scientific Conference on Sustainable, Modern and Safe Transport.
[7] L. Zhang, S. Dalyot, and M. Sester, Travel-Mode Classification for
Optimizing Vehicular Travel Route Planning, pp. 277–295. Berlin,
Heidelberg: Springer Berlin Heidelberg, 2013.
[8] H. Wang, G. Liu, J. Duan, and L. Zhang, “Detecting transportation
modes using deep neural network,” IEICE Transactions on Information
and Systems, vol. E100.D, no. 5, pp. 1132–1135, 2017.
[9] N. Breyer, D. GundlegÃ¥rd, and C. Rydergren, “Cellpath routing and
route traffic flow estimation based on cellular network data,” Journal of
Urban Technology, vol. 25, no. 2, pp. 85–104, 2018.
[10] Kalatian, Arash and Shafahi, Yousef, “Travel mode detection exploiting
cellular network data,” MATEC Web Conf., vol. 81, p. 03008, 2016.
[11] N. Breyer, D. GundlegÃ¥rd, and C. Rydergren, “Travel mode classi-
fication of intercity trips using cellular network data,” Transportation
Research Procedia, vol. 52, pp. 211–218, 2021. 23rd EURO Working
Group on Transportation Meeting, EWGT 2020, 16-18 September 2020,
Paphos, Cyprus.
[12] L. Novaˇ
cko, L. Šimunovi´
c, and D. Krasi´
c, “Estimation of
origin-destination trip matrices for small cities,” Promet -
Trafficamp;Transportation, vol. 26, pp. 419–428, Oct. 2014.
[13] H. Fanaee-T and J. Gama, “Event detection from traffic tensors: A hybrid
model,” Neurocomputing, vol. 203, pp. 22–33, 2016.
[14] M. Hubert and E. Vandervieren, “An adjusted boxplot for skewed
distributions,” Computational Statistics Data Analysis, vol. 52, no. 12,
pp. 5186 – 5201, 2008.
[15] S. Oh, Y.-J. Byon, and H. Yeo, “Improvement of search strategy
with k-nearest neighbors approach for traffic state prediction,” IEEE
Transactions on Intelligent Transportation Systems, vol. 17, no. 4,
pp. 1146–1156, 2016.
[16] D. Tong, Y. R. Qu, and V. K. Prasanna, “Accelerating decision tree
based traffic classification on fpga and multicore platforms,” IEEE
Transactions on Parallel and Distributed Systems, vol. 28, no. 11,
pp. 3046–3059, 2017.
[17] N. Dogru and A. Subasi, “Traffic accident detection using random forest
classifier,” in 2018 15th Learning and Technology Conference (L T),
pp. 40–45, 2018.
[18] S. Agarwal, P. Kachroo, and E. Regentova, “A hybrid model using
logistic regression and wavelet transformation to detect traffic incidents,”
IATSS Research, vol. 40, no. 1, pp. 56–63, 2016.
[19] J. Zhang, C. Chen, Y. Xiang, W. Zhou, and Y. Xiang, “Internet traffic
classification by aggregating correlated naive bayes predictions,” IEEE
Transactions on Information Forensics and Security, vol. 8, no. 1, pp. 5–
15, 2013.
[20] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Van-
derplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and Édouard
Duchesnay, “Scikit-learn: Machine learning in python,” Journal of
Machine Learning Research, vol. 12, no. 85, pp. 2825–2830, 2011.
[21] L. Tišljari´
c, D. Cvetek, and V. Vareški´
c, “Transport Mode Classification.”
https://github.com/tisljaricleo/transport-mode-classification, 2021.
63rd International Symposium ELMAR-2021, 13-15 September 2021, Zadar, Croatia
177
Authorized licensed use limited to: University of Zagreb: Faculty of Electrical Engineering and Computing. Downloaded on October 05,2021 at 11:35:13 UTC from IEEE Xplore. Restrictions apply.
