Received July 1, 2020, accepted July 30, 2020, date of publication August 7, 2020, date of current version August

19, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.3014901

Transportation Mode Detection by Embedded

Sensors Based on Ensemble Learning
BANDAR ALOTAIBI , (Member, IEEE)
Department of Information Technology, University of Tabuk, Tabuk 71491, Saudi Arabia
e-mail: [email protected]

ABSTRACT Context-aware computing has become a certainty due to the widespread use of smartphone
devices equipped with sensors. A wide range of services, such as vehicular traffic monitoring and smart
parking, can be accomplished with the help of awareness of user mobility. Transportation mode detection
(TMD) using machine learning algorithms and the data captured from smartphone embedded sensors
have attracted research community attention. In this research, ensemble learning is utilized to differentiate
between transportation modes, including walking, standing, riding a train, driving a car, and riding a bus.
The ensemble learning consists of three classifiers; each classifier votes independently on the instances,
and the majority vote is applied for robust generalization. The proposed method was validated using
three datasets; the samples included in these datasets were gathered by smartphone sensors (belonging
to heterogeneous users), such as rotation vector sensors, accelerometers, uncalibrated gyroscopes, linear
acceleration, orientation, speed, game rotation vector, sound, and gyroscopes. The proposed ensemble
learning method achieves an accuracy of 89%, 93%, and 95% on the first, second, and third datasets,
respectively, when 10% and 90% of the data are used for testing and training, respectively. On the other set of
experiments, in which 30% and 70% of the data are used for testing and training, respectively, the proposed
method yields accuracies of 86.8%, 92.1%, and 94.9% on the first, second, and third datasets, respectively.
The proposed method shows promising results compared to existing human activity recognition (HAR)
methods.

INDEX TERMS Context-aware computing, embedded sensors, ensemble learning, human activity recogni-
tion, Internet of Things (IoT), transportation mode detection.

I. INTRODUCTION ters and gyroscopes (a.k.a., microelectromechanical systems

Context-aware computing was introduced by Schilit and oth- (MEMS)) can be used to deduce what activity the phone
ers in 1994. The terminology is defined as ‘‘the possibility carrier is performing.
to exploit the changing environment with a new class of The awareness of a person’s transportation mode facilitates
applications that are aware of the context in which they are planning for urban transportation [7], which was performed in
run’’ [1]. Recently, context-aware computing has become a the past via telephone interviews, questionnaires, and travel
reality due to the ease of sensor deployment through IoT diaries [8], [9]. With these manual approaches, the traveler
devices (e.g., smart wearables such as smartwatches and could utilize several modes of transportation and might not
fitness trackers) or smartphone devices [2]. For instance, remember the certain times of the performed transportation
embedded sensors are integrated with smartphones to realize modes. Thus, context-aware computing replaces the tradi-
the surrounding areas, recognize user activity, and locate tional method, which is considered limited, not up-to-date,
user location [3]. A popular research topic (i.e., a type of erroneous, and expensive [10]. Moreover, human wellness
human activity recognition (HAR)) known as transportation and physical activities can be observed using smartphones
mode detection (TMD) has gained noticeable attention [4]– [11], smartwatches, or fitness trackers. Such applications
[6]. TMD is considered a HAR because smartphones carried can also provide daily activities and the number of calories
by users and embedded with sensors such as accelerome- burned. Additionally, the activities that affect the environment
can also be estimated, and environmental hazard exposure
The associate editor coordinating the review of this manuscript and can be traced. Knowing the transportation mode of travelers
approving it for publication was Giovanni Pau . helps shops send real-time information to customers in the

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
145552 VOLUME 8, 2020
B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

form of advertisements. For instance, detecting a consumer extreme gradient boosting (XGBoost), have been investigated
driving a car allows for gas stations to send vehicle ser- to differentiate the mode of transportation. However, these
vices or gas coupons as advertisements. Unlike life critical ensemble methods can only combine machine learning algo-
systems that cause death or injuries to people, environmental rithms with identical types, such as decision trees (DTs).
hazards, or expensive equipment damage, such as heart-lung However, ensemble methods that combine machine learning
machines, ventilation systems, and fire alarms, where an ideal algorithms with different types, such as stacked learning,
100% precision is essential, transportation mode detection have not been utilized by the TMD research community.
can tolerate a margin of error ranging from 5% to 10%. Stacked learning bundles lower level models of different
Currently, three distinguished types of methods are uti- types by a higher level generalized model to maximize accu-
lized to detect transportation modes: sensor-based techniques racy.
[12], global positioning system (GPS)-based techniques [13], Motivated by the released public datasets and the room
and cellular-network-based attempts. Furthermore, several for improvement, a stacked learning model consisting of
research papers [14]–[16] have attempted to build patterns of various machine learning algorithms with different types (i.e.,
transportation modes and improve the accuracy of these tech- the investigated machine learning algorithms in this study
niques. Moreover, some industries, through their applications that generate the most accurate final prediction when bundled
and operating systems (e.g., Android operating system), have together in the stack) is presented to detect the transportation
implemented their own TMD in which users are informed mode using available smartphone sensors that remarkably
of their mode of transportation. However, some of these improve the detection accuracy compared to TMD methods.
applications or systems limit their capabilities of detecting The contributions of this research can be summarized as
activities to four categories: riding a bicycle, riding in a follows:
vehicle, standing, and walking [3]. All of the techniques share
the following identical general principle for the detection of 1) A comprehensive study of different machine learning
the transportation mode: collecting raw data from sensors algorithms is presented to investigate which machine
integrated with mobile phones, preprocessing the raw data learning algorithm best fits the problem of TMD.
to be fed into the classification model, feeding part of the 2) Based on the observations from this study, a new
preprocessed data into the classifier for training purposes, and ensemble method is proposed to differentiate trans-
passing the rest of the preprocessed data to the predictor for portation modes. This method achieves promising
generalization purposes. results compared to existing techniques.
Although TMD methods have achieved considerable suc-
The remainder of this article is organized as follows.
cess in many applications, there are still some limitations
Section II summarizes the related methods, the proposed
regarding the studies that have been proposed to help improve
method is introduced in Section III, and IV presents the
the usefulness of these methods. First, due to the lack of a
experimental settings and the utilized dataset that help to eval-
benchmark dataset, the published results are hard to compare.
uate the proposed method. Section V presents the parameter
Additionally, the proposed techniques are not well defined
optimization, Section VI validates the proposed method and
because of either the limited number of smartphone users
analyzes the results, Section VII discusses the results, and
involved in the testbed or the heterogeneity of sensors that
Section VIII concludes the research paper.
have been utilized (i.e., it is hard for industries to adopt
one of these methods). Fortunately, two recent datasets [3],
[17] have been published in which researchers can compare II. RELATED WORK
their results. Moreover, a large number of proposals have Due to the widespread use of smartphones and the Internet
used binary classification in which the classifier needs to of things (IoT) devices that are equipped with embedded sen-
distinguish between only two modes of transportation. For sors, the context-aware computing topic has regained inter-
instance, smart parking techniques only considered modes est in the research community. TMD is an important field
of transportation: riding a vehicle and walking. The notifi- in which embedded sensors play an integral role in fulfill-
cation is sent once a busy or free area is found [18]. Finally, ing its purpose. TMDs can be categorized into three main
various techniques have been proposed to detect the mode types of techniques: sensor-based, GPS-based techniques,
of transportation using a single machine learning algorithm and cellular-network-based attempts. GPS-based techniques
to improve the accuracy. Certain sets of samples might be are the most famous category because of the ease of acces-
fit well by a single machine learning algorithm; however, sibility to GPS in smartphones and the high achieved perfor-
overfitting or underfitting might become visible to the rest mance in terms of accuracy in differentiating between pedes-
of the samples. Thus, even with parameter optimization, trian movements and riding vehicles. The GPS-based tech-
an upper limit of accuracy might be reached when utilizing niques share two shortcomings: the accuracy degrades once
one machine learning algorithm. Combining various machine the smartphone carrier enters an indoor environment or passes
learning algorithms (ensemble learning) is used to overcome by an urban canyon due to multipath fading and smart-
this shortcoming [19]. Recently, several ensemble methods, phones equipped with battery-constrained capability (i.e.,
such as random forests (RFs), gradient boosting (GB), and these methods consume batteries) [15].

VOLUME 8, 2020 145553

 B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

Cellular network-based attempts employ record-keeping (i.e., to combine important features that could improve the
via three techniques to determine the transportation mode performance) and RFs to improve the performance.
of mobile phone owners: signaling data, handovers, and call Xiao et al. [25] proposed a TMD method based on ensem-
detail records (CDRs). These records are kept by telecommu- ble learning in which three machine learning algorithms were
nication companies for management assistance, maintenance, combined. The authors used the GPS to detect the trans-
and billing. Consequently, users are not required to make any portation mode. To create global features and extract various
efforts and are not aware of the stored records that can be used local features, the authors utilized a statistical technique in
to estimate their locations and mode of transportation through which features were generated from subtrajectories. Then,
cellular towers. These methods share a deficiency related to these features were fed into the classifier in the training
revealing the data privacy of end users. Although mobile phase. The ensemble model consists of three base methods:
phone networks employ anonymity mechanisms to conceal GB, XGBoost, and RFs. The most accurate method among
the identity or location of end users before data analysis, the base estimators is the XGBoost method (its accuracy is
recognizing four spatiotemporal points can be a key factor 90.77%) when it is applied to the GEOLIFE dataset [26]. The
in revealing the identity or location of a given user with high authors also reduced the time required to predict the mode of
probability. Thus, new techniques, i.e., those compatible with transportation by using a feature selection method (i.e., a tree-
governmental authorities or intergovernmental organizations based ensemble technique).
(e.g., the European Union (EU)), and regulations such as The authors in [27] studied transportation mode classifi-
general data protection regulations (GDPRs) should be devel- cation utilizing big data generated by smartphone embedded
oped to protect the data privacy and anonymity of mobile gyroscope, magnetometer, and accelerometer sensors. They
phone users [20]. investigated three well-known machine learning algorithms
Sensor-based methods use the raw data captured by smart- to recognize transportation modes, specifically, SVMs,
phone embedded sensors. The captured data in the form of KNNs, and DTs. The authors used large-scale experiments
series reading are preprocessed, and the dimensionality of in which the performance of these methods was compared in
the captured data might be reduced to improve the perfor- terms of accuracy, model size, and prediction time. The per-
mance or to make the approach less complex. The prepro- formed experiments show promising results in which SVMs
cessed data are then fed into a machine learning algorithm perform better than the other two methods in terms of accu-
for training purposes. Consequently, the weight of the training racy, but it was the most expensive method in terms of time.
phase is used to predict the mode of transportation captured Shafique and Hato [28] examined four machine learn-
by the smartphone embedded sensor. The most likely smart- ing algorithms, specifically adaptive boosting (AdaBoost),
phone embedded sensor that best fits the area of the TMD RFs, DT, and SVM, to differentiate the exercised mode of
is the accelerometer, which has two advantages: its ability to transportation. There are four detected modes of transporta-
help detect human activities and its light energy consumption tion: walk, train, bike, and car. The generated data are col-
[21]. One of the most widely used machine learning algo- lected from three Japanese cities, and the target sensor is an
rithms in TDM is RFs [22]–[24]. accelerometer. Among the tested methods, RFs outperform
The authors of [4] investigated several machine learning the other methods. The second best method is DT, then SVMs
algorithms to generalize a multiclass method capable of rec- and the worst method in terms of accuracy is AdaBoost.
ognizing the mode of transportation. The dataset that the Fang et al. [29] proposed a deep learning-based approach
authors used consists of five classes: bike, car, bus, walk, to effectively construct a nonlinear shape between labeled and
run. The authors tested various machine learning methods, sequential data captured by sensors. This approach utilized a
including support vector machines (SVMs), DT, k-nearest deep neural network (DNN) and built two types of feature
neighbors (KNNs), bagging, and RFs. The raw data were cap- sets. One feature set was small compared to the other feature
tured from three smartphone embedded sensors: the rotation set to investigate the tradeoff between performance and bat-
vector, accelerometer, and gyroscope. During the training and tery consumption. The other only considered the performance
testing phases, the authors utilized both the out-of-bag error of the proposed method. The authors compared their method
and k-fold cross-validation to select the model that performs with three machine learning algorithms: SVMs, KNNs, and
better. A feature selection method known as minimum redun- DT. The proposed method yielded better results compared to
dancy maximum relevance (mRMR) was used to eliminate the other three methods.
the features that negatively affect the classifier performance. Liang and Wang [30] proposed a deep learning tech-
The authors found that data captured by smartphone embed- nique (i.e., convolutional neural network (CNN)) to automat-
ded sensors provide meaningful patterns that distinguish ically detect the transportation mode using data generated by
between different types of transportation modes. The authors accelerometer sensors. Various machine learning algorithms
compared the performance of the tested machine learning are tested to compare the proposed method with the other
algorithms after applying a feature selection method and methods. The authors tried several CNN architectures and
found that the RFs ensemble method and SVMs performed proposed a lightweight and accurate architecture. The pro-
better than the other methods. The authors also attempted posed architecture is compared with several machine learn-
to combine both the simulated annealing (SA) algorithm ing algorithms. The proposed CNN model outperforms the

145554 VOLUME 8, 2020

B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

other methods, and Gaussian naive Bayes (GNB) is the least activity, thus:
accurate method among the tested methods. 


 0 if s represents standing still activity

1 if s represents riding a bus activity


III. PROPOSED METHOD t= 2 if s represents riding in a vehicle activity
The proposed framework shown in Fig. 1 utilizes several 
3 if s represents riding a train activity

machine learning algorithms to discriminate the mode of



4
transportation performed by the subjects. The predictions of if s represents a walking activity
three machine learning algorithms namely, DT [31], RFs [32], The whole training set can be denoted by:
and GB [33], are fed into an ensemble to better generalize
the predictions by taking the majority vote of these three d = (s1 , t1 ), (s2 , t2 ), . . . , (sn , tn ) (1)
algorithms. Subsequently, the predictions of the ensemble Moreover, the feature space F (i.e., always the vertical
along with two other machine learning algorithms, namely, depiction of the dataset) consists of several characteristics
KNNs [34] and random subspace [35] (a.k.a., bagging) are representing smartphone embedded sensor measurements
bundled into stacked learning in which a neural network applied to all the samples in the dataset. For instance, the first
architecture known as multilayer perceptron (MLP) [36] is feature f1 in the dataset can represent the average of the
used to eventually predict the mode of transportation. Twelve first sensor measurement, the second feature f2 the standard
machine learning algorithms are intensively tested using two deviation of the first sensor measurement, and so on.
data splits to determine which machine learning algorithms The proposed model is divided into three phases. In the
when bundled together in the stack produce a final prediction first phase, the samples and their labels are assigned to three
that best fits the TMD problem. Thus, the base estimators classifiers: DT, GB, and RFs. Given both the sample vectors
are selected because when bundled together in the stack, in the form of real numbers: Sn ∈ Ri , where n = 1, . . . , N
the final prediction of the stack is the most accurate. With the represent the samples and the target vector T ∈ Rj , the DT
exception of SVMs that are eliminated because of the time classifier groups the instances with identical labels together
complexity (i.e., linear SVM (LSVM) and nonlinear SVM by recursively partitioning the dimension space. Since it is a
(NLSVM)), these machine learning algorithms, including multiclass classification problem, the output is in the form of
RFs, bagging, GB, XGB, AdaBoost [37], DT, MLP, KNNs, 0, 1, . . . L − 1, for a given decision node d, exemplifying a
logistic regression (LR), and GNB are tested in each phase of region Rd in the dimension space in which the observation is
the proposed method in an interchangeable manner, and the Zd , thus
most accurate attempt is chosen.
1 X
rdl = I (Tn = l) (2)
Zd
Sn ∈Rd

is the probability of l observations in the decision node d,

where I (.) is an indicator function that receives 1 if the
statement is true, and 0 otherwise; therefore, the value of
I (Tn = l) is 1 if this statement is true, and 0 otherwise.
Then, the Gini impurity function G is applied
X
G(Xd ) = rdl (1 − rdl ) (3)
l
where Xd represents the training data in decision node d.
In that phase (i.e., the first phase), the samples and their
labels are given to another machine learning technique,
namely, GB, in which base estimators (a.k.a., weak learners),
usually in the context of DTs, are the building blocks that
fulfill the demand of the final substantial ensemble prediction
model. DTs grow sequentially with respect to errors. Each
FIGURE 1. Representation of the proposed stacked method that is tree learns from the mistakes of the previous tree to improve
bundled with three machine learning algorithms, namely, KNNs, bagging, performance. Let M1 (s) = t denote the DT model that fits
and an ensemble method consisting of three base estimators: DT, GB, and
RFs. the data. The next DT H1 (s) learns from the mistakes of the
previous DT
H1 (s) = t − M1 (s) (4)
More formally, to train the model, let s represent the This new DT is added to GB, which is denoted as
activity measurements captured by the subject smartphone
embedded sensors and t denote the target class labeling that M2 (s) = M1 (s) + H1 (s) (5)

VOLUME 8, 2020 145555

 B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

This procedure continues until interruption occurs by a where δ is a delta function known as the Kronecker delta
certain mechanism such as cross validation. The ensemble named after its inventor Leopold Kronecker.
model, known as an additive model of o individual trees, The last technique in this phase is bagging (a.k.a., ran-
is built after several stages are performed (similar to other dom subspace). Different bootstrap samples are generated
boosting methods), as shown in the next equation, and an arbi- by invoking a base learning algorithm to train several base
trary differential loss function is optimized for generalization learners. Subsampling the data with replacement in the train-
purposes [38]. ing set produces bootstrap samples that have the same size
O
as the size of the training set. Once the base learners are
m(s) =
X
mo (s) (6) acquired, the random subspace technique is used to combine
the base learners and takes the majority vote in which the
o=1
prediction occurs when a specific class gets the majority.
The third machine learning algorithm utilized in this phase More formally, let B denote the base learning algorithm, and
is RFs. RFs is an ensemble method consisting of trees, and R is the number of rounds in the learning process. For all the
each tree utilizes random variable collection. Given the sam- rounds r = 1 . . . , R, bootstrap samples bs are produced from
ples and their labels to RFs, the aim is to predict the class the data d (i.e., dr = bs (d)), and a base learner hr is trained
target t using a prediction function f (S) in which this function from the bootstrap samples hr = B(dr ). Thus, the iteration of
is used to reduce the loss value and defined by a given this process generated the following output.
loss function. The prediction function is constructed utilizing
R
some base estimators e1 (s), . . . , ei (s). The base estimators X
f (s) = arg max I (t = hr (s)) (11)
are integrated to establish the ensemble predictor f (s) using t∈T 0
r=1
voting as follows:
where the value of I (t = hr (s) is 1 if this statement is true
I
X and 0 otherwise [41].
f (s) = arg max I (t = ei (s)) (7) In the last phase, a metalearner (i.e., a special type of DNN
t∈T 0
i=1 called MLP) is utilized to combine the second phase learning
where T 0 is a set of potential values of t. algorithms, namely, ensemble learning, KNNs, and bagging,
In the second phase, the predictions of the three classifiers that have been produced from the training data. At least
are passedto the th three layers are involved in constructing MLP: an input layer,
ensemble. Let the classifier’s decision of c
be Dc,t ∈ 0, 1 . where c = 1 . . . , C represent the classifiers. a hidden layer, and an input layer. The hidden layer is the
If the target class t0 is chosen by the cth classifier, then dc,t = computational unit (neuron). The objective of the hidden
1 and 0 otherwise. The plurality voting is utilized in which the layer is to receive an input from other sources or hidden layers
predicted class is the one that has received the highest number and generate an output. The following equation is utilized by
of votes, as shown in the next equation [39]: MLP:
C
X C
X f (s) = V2 j(V1T s + a1 ) + a2 (12)
Dc,t ∗ = max Dc,t (8)
c where V1 is the input layer weight, V2 is the hidden layer
c=1 c=1
weight, a1 is the hidden layer bias, a2 is the output layer
Two other machine learning techniques come into play bias, and j is the activation function. This research utilizes
in this phase: KNNs and bagging. The input of the KNNs the hyperbolic tangent activation function, which can be rep-
algorithm is the k-closest training instances constructed in the resented as:
feature vector. The output is a member of a specific class in yx − y−x
which a given object is selected by the neighbors of that object j(x) = x (13)
y + y−x
using a plurality vote. The function of KNNs is approximated
locally, and the classification is carried out after postponing where yx −y−x represents the hyperbolic sine of y and yx +y−x
all computations; thus, KNNs is a typical example of lazy represents the hyperbolic cosine of y.
learning or instance-based learning [40]. Let K (s) = t denote f (s) is the size of a vector of the number of classes that
the label function in which a training instance is assigned a passes via the softmax function as follows:
class target out of l possible labels t ∈ 0, 1 . . . , L − 1 by that exp(xn )
function. Thus, the nearest neighbors of k for a query data softmax(x)n = Pk (14)
i=1 exp(xi )
point s[p] are:
The nth element of input to softmax is denoted by xn ,
Qk = s[1] , K s[1] . . . , s[k] , K s[k]





(9) and the number of classes is represented by k. This function
produces a vector consisting of the probabilities in which
and the hypothesis of the KNNs can be defined as: sample s is categorized to each class and the class with the
k
X highest probability is the output [42].
h s[p] = arg  max δ(t, K s[i] )



(10) The dataset consisting of the samples and their labels is
the input to this phase. Let B1 . . . , BR denote the second

t∈ 0,1...,L−1 i=1

145556 VOLUME 8, 2020

B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

phase learning algorithms and B denote the current phase TABLE 1. Dataset consisting of sensor measurements collected from
thirteen different subjects with diverse genders, occupations, device
learning algorithm. For all the rounds r = 1 . . . , R, a second models, and ages.
phase individual learner dr using the second phase learning
algorithm br and applying it to the original dataset d, so math-
ematically speaking dr = bs (d). A new dataset d 0 = ∅ is
generated from the previous process. For all the new training
samples n = 1 . . . , i and for all the new rounds r = 1 . . . , R,
classify the training instance si by utilizing dr (i.e., znr ) to
produce the following dataset.
d0 = d0 ∪ zn1 , zn2 , . . . , znR , tn



(15)
After the new dataset is completely generated, the current
phase learner c0 is trained utilizing the current phase learning
algorithm B and the newly produced dataset d0 ; thus, c0 =
B(d 0 ). This phase generates the following output:
f (s) = c0 c1 s . . . , cR s




(16) activities, enter their names, and label the performed activ-
ities. Some information is provided once the change in the
IV. EXPERIMENTAL SETTINGS measured parameters occurs, such as the sensor’s name, sen-
The device that is used to evaluate the proposed method sor’s raw data, and timestamp. Each measurement is stored
is a laptop equipped with a Windows 10 operating system, in a comma-separated values(CSV) file; thus, the dataset
i7 central processing unit (CPU), and 32 gigabyte (GB) ran- contains 226 CSV files (i.e., a total of 31 hours of recording,
dom access memory (RAM). The Anaconda distribution and as shown in Table 2) representing the actions taken by the
Python programming language are used to utilize the well- individuals performing the five activities. The distribution of
known machine learning library (i.e., Scikit-learn). these activities in the dataset is as follows: walking activities
To validate the generalization of the proposed method, correspond to 26% of the dataset, driving car activities cor-
a recent dataset [43] is used, which is one of the only two respond to 25%, standing still class records represent 24%,
TMD sensor-based public datasets that have been found. This riding a train represents 20% of the dataset, and only 5%
dataset consists of 5,894 trips representing five modes of represents measurements of individuals riding a bus.
transportation. As shown in Table 1, this dataset is favored
because it is more thorough, unlike the other dataset [44] TABLE 2. A total of 31 hours of recording, representing the actions taken
that captures sensor measurements from three users using by the individuals performing the five activities.

one device model (i.e., Huawei Mate 9). This dataset col-
lects sensor measurements from thirteen different users (three
female and ten male) with diverse genders, occupations,
Android versions, device models, and ages using the Android
app. There are no restrictions imposed by the creators of the
dataset on the use of the Android app that captures the mode The collected dataset is a combination of measurements
of transportation. Therefore, each user records the sensor captured by all available twenty-three sensors in smart-
measurements when an activity is performed to simulate real- phones. Among these twenty-three available sensors, fifteen
world situations. Each user assigns an action to perform to sensors are supported by enough smartphones, so these sensor
record the sensor measurements generated by that action. The measurements are retained for further analysis. Due to bias
modes performed by these individuals include standing still, measurements captured by some sensors, six sensors were
walking, riding in a car, riding a bus, and riding a train. The eliminated from the dataset, including pressure, magnetic
dataset has been constructed to follow general practices taken field, uncalibrated magnetic field, proximity, gravity, and
by authors of other datasets in the TMD area [12]. light. The remaining nine sensors are used to construct the
The sensor events are then gathered by an application dataset, including rotation vector sensors, game rotation vec-
that sets a maximum frequency of 20 hertz (Hz). The max- tor, accelerometers, uncalibrated gyroscopes, linear acceler-
imum sampling rate of 20 Hz is a decent choice because ation, orientation, speed, sound, and gyroscopes. Measure-
the frequency elements of body motions are measured and ments generated by speed and sound sensors are converted to
recognized to be lower than 20 Hz [45], [46]. Furthermore, positive values.
the vibration frequency of the vehicle seat while the engine is The dataset was divided into time windows. The size of
on ranges between 3 and 5 Hz, even when the vehicle is not in the time window was set to 5 seconds by the creators of the
motion [47]. The subjects using this application were asked dataset for two reasons: to capture the action being performed
to perform the five activities and were authorized to start and and following other researchers’ suggestions [12], [48]–[50]
stop capturing measurements when performing one of these in which a 5-second window was the perfect candidate to

VOLUME 8, 2020 145557

 B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

TABLE 3. The first dataset contins 12 features, the second dataset encompasses 32 features, and the third dataset comprises 36 features. Each dataset
consists of 5,894 samples which are measured by smartphones carried by 13 subjects, accumulated of 31 hours of recording, and sampled at a maximum
frequency of 20 Hz. Note: maximum is denoted by max, min denote minimum, µ denote mean, and standard deviation is denoted by σ .

measure the whole action being performed by a specific sub- their models, the default values such as 2 minimum number
ject. Therefore, setting a small window size results in misclas- of samples split, 1 minimum number of sample leafs, and
sification because some modes require a certain amount of Gini criterion function (i.e., responsible for measuring the
time that is greater than 3 seconds (e.g., to reveal the activity split’s quality) were utilized. The important hyperparame-
characteristics of walking, at least 3 seconds of accumulated ters of the other base estimators were chosen based on grid
data is needed). Correspondingly, setting a large window size search. Among (10, 20, 30, 40, 50, 60, 70, 80, 90, and 100)
introduces noise from various other activities interfering with trees in RFs, GB, and bagging, 10, 100, and 10 were selected,
the targeted mode [45]. For the nine sensors, the maximum, respectively.
the minimum, the mean, and the standard deviation have In the second phase of the proposed method, grid search
been calculated for each window size and inserted into three utility was utilized to select the most valuable KNNs algo-
datasets that represent the features (i.e., columns in the CSV rithm options to determine the mode of transportation from
files). Thus, for each sample (i.e., 5-second window size), three hyperparameters, namely, metrics of similarity, number
a total of 4 features for each sensor was constructed. More- of neighbors, and the weight function. From three metrics
over, the missing values were averaged with the values of the of similarity, namely, Euclidean, Manhattan, and Minkowski,
given feature. the latest was chosen. Five neighbors were investigated (3, 5,
Furthermore, since the number of samples measured by the 7, 9, and 11) to optimize this parameter, in which 3 was the
sensors representing bus activity is much lower than those of best fit for the data. The weight function was the last explored
the rest of the activities, as shown in Table 2, all the mea- parameter with two options, namely, uniform or distance,
surements have been lowered to match the bus measurements in which uniform was the chosen function. The purpose of
(i.e., to balance the samples belonging to each class). The the uniform function is to equalize the weight for the data
sensor measurements construct three benchmark datasets for points in each neighborhood, while the distance function
evaluation purposes. The number of samples for the dataset is gives closer neighbors of the query data point more influence
5,894, representing the activities performed (i.e., the journeys than the rest of the data points.
taken) by the 13 subjects. As shown in Table 3, the first In the last phase of the proposed method, four MLP hyper-
dataset consists of 12 features in the form of maximum, parameters were considered, namely, the number of hidden
minimum, mean, and standard deviation calculations, exem- layers, the number of neurons in each hidden layer, the activa-
plifying the measurements captured by three sensors, namely, tion function, and the optimizer. In addition to one input layer
the gyroscope, accelerometer, and sound. The second dataset and one output layer, two hidden layers were used to mini-
contains 32 features representing the measurements captured mize the time complexity. The first layer consists of 80 neu-
by the rest of the sensors, with the exception of speed (i.e., rons, while the second layer consists of 10 neurons. For
the gyroscope, uncalibrated gyroscope, game rotation vector, the hidden layers, identity, logistic, hyperbolic tangent (i.e.,
rotation vector, linear acceleration, orientation, accelerom- tanh), and ReLU activation functions were tested, in which
eter, and sound). The last dataset consists of 36 features hyperbolic tangent performed better than the other three func-
representing all the sensor measurement data, including those tions. From the two tested weight optimizers (Adam and
captured by the speed sensor. stochastic gradient descent), Adam was chosen because it
works very well with datasets that have a sufficient number of
samples (i.e., thousands of samples) such as the TMD dataset.
V. PARAMETER OPTIMIZATION
To improve the performance of the proposed method in terms
of the four evaluation measures while maintaining decent VI. RESULTS
time complexity, the parameters of the base estimators were This section presents the results of the proposed method along
optimized. To reduce the complexity of the ensemble methods with the results of the different machine learning algorithms
such as RFs, GB, and bagging that rely on DT to build that have been tested in the experiments. The performance

145558 VOLUME 8, 2020

B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

TABLE 4. Evaluation measures of all algorithms on the three datasets using two different data splits sorted by accuracy in descending order.

of the tested methods along with the proposed technique the proposed method was better than the second-best method
were evaluated using four measures: precision, recall, f1- (i.e., GB) by approximately 3%, 2%, and 2% on the first,
score, and accuracy. The precision is calculated as FPTP + TP , second, and third datasets, respectively.
TP The details of each correctly classified class of samples and
the recall is calculated as TP + FN , the f1-score is calculated
precision · recall
as 2 · precision misclassified samples are shown in Table 5 in the form of a
+ recall , and the accuracy is calculated as
TP + TN
confusion matrix. In the first set of experiments (when the
TP + TN + FP + FN . first dataset was split into 90% training and 10% testing),
In the first set of experiments, each of the three datasets the still class samples were correctly classified 96% of the
was split into 90% training and 10% testing. The proposed time, the bus class samples were correctly classified approxi-
method was evaluated and compared with the other machine mately 91% of the time, the car class samples were correctly
algorithms using four evaluation metrics, namely, accuracy, classified approximately 78% of the time, the train class
precision, recall, and F1-score, as shown in Table 4. The samples were correctly classified approximately 90% of the
proposed method outperformed the rest of the methods when time, and the walking class samples were correctly classified
tested on the first dataset; its accuracy is 89%, which is better 91.5% of the time. In the second dataset (using the same data
than the second-best method (i.e., RFs) by approximately 3%. split, i.e., 90% training and 10% testing), the still and bus
It is also better than the other methods when tested on the sec- class samples were correctly classified approximately 97%
ond dataset; its accuracy is 93.2%, which is better than the of the time, the car class samples were correctly classified
second-best method by 2.4%. Finally, the performance of the approximately 90% of the time, the train class instances were
proposed method is similar to that of the second-best method correctly classified 91% of the time, and the walking class
on the third dataset; it is better than the second-best method samples were correctly classified 92% of the time. In the third
by only 0.5%. In the second set of experiments, each of the dataset, the still class samples were correctly classified 98%
three datasets was split into 70% training and 30% testing. of the time, the bus class samples were correctly classified
As shown in Table 4, the proposed method outperformed the approximately 97% of the time, the car class samples were
rest of the methods on the three datasets. The accuracy of correctly classified approximately 93% of the time, the train

VOLUME 8, 2020 145559

 B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

TABLE 5. Confusion matrices for the stacked proposed method.

class samples were correctly classified approximately 95% second, and third datasets, respectively. The testing times of
of the time, and the walking class samples were correctly the proposed method using the three datasets were 4 ms for
classified 95% of the time. the three datasets.
In the second set of experiments (when the first dataset
was split into 70% training and 30% testing), the still class VII. DISCUSSION
samples were correctly classified 94% of the time, the bus The accuracy of our method outperformed the other methods
class samples were correctly classified approximately 88% on the three datasets using two data splits (i.e., 70% training–
of the time, the car class samples were correctly classified 30% testing data split and 90% training–10% testing data
approximately 78% of the time, the train class samples were split). The second-best method in terms of accuracy on the
correctly classified approximately 82% of the time, and the first dataset was RFs. The second-best method on the second
walking class samples were correctly classified 93% of the dataset was GB. The effect of the data split and sensors
time. In the second dataset (using the same data split, i.e., 70% involved in gathering sensor measurements from smartphone
training and 30% testing), the still class samples were cor- devices equipped with sensors carried by the subjects in
rectly classified approximately 94% of the time, the bus class experiments is shown in Fig. 2.
samples were correctly classified approximately 97% of the
time, the car class samples were correctly classified approxi-
mately 87% of the time, the train class samples were correctly
classified 90% of the time, and walking class samples were
correctly classified 93% of the time. In the third dataset,
the still class samples were correctly classified 97% of the
time, the bus and train class samples have been correctly clas-
sified approximately 96% of the time, the car class samples
were correctly classified approximately 90% of the time, and
the walking class samples were correctly classified 95% of
the time.
The training and testing times are shown in Table 6. The
training time of the proposed method was adequate com-
pared to the other methods, while the testing time was lower
compared to the rest of the methods. In the first set of
experiments where the dataset was split into 90% training
and 10% testing, the training time of the proposed method
using the first dataset was approximately 4.4 s, approximately
FIGURE 2. Two data splits were used to measure the performance of the
4.5 s for the second dataset, and approximately 4.4 s for the algorithms. Moreover, the effect of the sensors involved in the three
third dataset. The testing time of the proposed method was datasets was investigated.
1 ms when applied to the first dataset, 2 ms when applied to
the second dataset, and 2 ms when applied to the third dataset. It is obvious from the figure that as the training data
In the second set of experiments where the dataset was split increased (increasing the training from 70% to 90%),
into 70% training and 30% testing, the training times of the the accuracy of all the algorithms increased as well. More-
proposed method were 3.2 s, 3.3 s, and 3.5 s on the first, over, the number of sensors integrated with smartphones

145560 VOLUME 8, 2020

B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

TABLE 6. Training and testing times of all algorithms on the three datasets using two different data splits.

affects the performance. As shown in the figure, the first TABLE 7. The ten most important features sorted by importance in
descending order.
dataset contains measurements captured by three sensors:
sound, accelerometer, and gyroscope. Due to the limited
sensors involved in the dataset, the performance of all the
algorithms did not exceed 90% (the most accurate algo-
rithm was the proposed method with 90% accuracy when
applied to a 10% testing data split). As the number of sensors
increased, the results of all the algorithms increased as can be
seen when the second dataset was used, which consisted of
measurements captured by an accelerometer, game rotation
vector, gyroscope, gyroscope uncalibrated, linear accelera-
tion, orientation, rotation vector, and sound. The most accu-
rate technique was the proposed method (its accuracy was
93.2% when applied to a 10% testing data split). Additionally, time for one measurement was approximately 7 microseconds
the accuracy of all the algorithms increased when the remain- (i.e., the result of the division of the testing time and the
ing sensor measurements (i.e., speed sensor) were involved number of measurements in the test set).
in the third dataset. The best method remained the proposed Obviously, the accuracy of transportation mode detection
method with an accuracy of 95.6%. increases as the measurements of more sensors are involved in
Since RFs is the second-most accurate method in the first the detection process. However, increasing the number of sen-
set of experiments (after the proposed method) and the third- sors might increase the complexity of the detection method
most accurate method in the second set of experiments, (i.e., the time required to detect the transportation mode) and
the feature importance capability provided by it was utilized raise some privacy concerns. Although the proposed method
to conceive the features that are considered most important in achieved superior results compared to the tested machine
distinguishing the mode of transportation. The second set of learning methods and can detect the mode of transportation
experiments was used to screen the most important features, in a reasonable time, there is room for improvement, which
where the data were split into 70% training and 30% testing. could be investigated further in future work. A variety of
Because the third dataset contains all the features (i.e., a total lower sampling rates, such as 1 Hz, could be utilized in future
of 36 features), it was used. The ten most important features in work, as suggested by some GPS-based techniques [51] to
distinguishing the mode of transportation applied to the third improve the detection rate of transpiration mode. In addition,
dataset on the second set of experiments are shown in Table 7. the geographic zones taken by participants while performing
The training time and testing time of all the algorithms the transportation mode should be recorded and included in
were decent. Although the training time of our algorithm was future work to clarify the route taken by participants. More-
not the best compared to the other algorithms, the training over, various window sizes should be studied in future work
time was not as important as the testing time because the (e.g., a window size of 1 minute has been suggested by some
training was performed only once. The testing time of the GPS-based approaches [?], [52]).
proposed method was short. The number of measurements
was 5,894, and the testing time for 10% (i.e., approximately VIII. CONCLUSION
589 measurements) of the data was 4 ms when the proposed Context-aware computing is a significant area that has been
method was applied to the third dataset, so the prediction accomplished in the last two decades with the help of

VOLUME 8, 2020 145561

 B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

smartphone devices equipped with sensors. A variety of [10] P. Widhalm, P. Nitsche, and N. Brändie, ‘‘Transport mode detection with
services ranging from smart parking to vehicular traffic mon- realistic smartphone sensor data,’’ in Proc. 21st Int. Conf. Pattern Recognit.
(ICPR), Piscataway, NJ, USA, 2012, pp. 573–576.
itoring can be achieved utilizing the awareness of user mobil- [11] K. Lee and M.-P. Kwan, ‘‘Physical activity classification in free-living con-
ity. Detecting the mode of transportation with the help of ditions using smartphone accelerometer data and exploration of predicted
machine learning has gained researchers’ attention due to the results,’’ Comput., Environ. Urban Syst., vol. 67, pp. 124–131, Jan. 2018.
[12] L. Bedogni, M. Di Felice, and L. Bononi, ‘‘Context-aware Android appli-
various services that can help accomplish this. In this article, cations through transportation mode detection techniques,’’ Wireless Com-
a stacked learning technique was proposed to detect the mode mun. Mobile Comput., vol. 16, no. 16, pp. 2523–2541, Nov. 2016.
of transportation, such as riding a train, driving a car, stand- [13] S. Reddym, M. Mun, J. Burke, D. Estrin, M. Hansen, and M. Srivastava,
‘‘Using mobile phones to determine transportation modes,’’ ACM Trans.
ing, walking, and riding a bus, utilizing three datasets that Sensor Netw., vol. 6, no. 2, p. 13, Mar. 2010.
were constructed to measure the activity performed by differ- [14] T. Feng and H. J. P. Timmermans, ‘‘Transportation mode recognition using
ent subjects using smartphones equipped with various sen- GPS and accelerometer data,’’ Transp. Res. C, Emerg. Technol., vol. 37,
pp. 118–130, Dec. 2013.
sors. Extensive sets of experiments that involve 12 machine
[15] X. Su, H. Caceres, H. Tong, and Q. He, ‘‘Online travel mode identification
learning algorithms were conducted to validate the proposed using smartphones with battery saving considerations,’’ IEEE Trans. Intell.
method. The proposed method outperformed the rest of the Transp. Syst., vol. 17, no. 10, pp. 2921–2934, Oct. 2016.
methods in terms of accuracy on the three datasets with two [16] S. Bhattacharya and N. D. Lane, ‘‘From smart to deep: Robust activity
recognition on smartwatches using deep learning,’’ in Proc. IEEE Int. Conf.
different data splits. The prediction time of each measurement Pervas. Comput. Commun. Workshops (PerCom Workshops), Piscataway,
was so low that it is an ideal option to detect the mode of NJ, USA, Mar. 2016, pp. 1–6.
transportation in real time. In the first set of experiments [17] L. Wang, H. Gjoreski, M. Ciliberto, S. Mekki, S. Valentin, and D. Roggen,
‘‘Enabling reproducible research in sensor-based transportation mode
where the dataset was split into 10% testing and 90% training, recognition with the Sussex-Huawei dataset,’’ IEEE Access, vol. 7,
the proposed method yielded accuracies of 89%, 93%, and pp. 10870–10891, 2019.
95% on the first, second, and third datasets, respectively. [18] J.-G. Krieg, G. Jakllari, H. Toma, and A.-L. Beylot, ‘‘Unlocking the
smartphone’s sensors for smart city parking,’’ Pervas. Mobile Comput.,
In the second set of experiments where the dataset was split vol. 43, pp. 78–95, Jan. 2018.
into 30% testing and 70% training, the proposed technique [19] Z. Ma, P. Wang, Z. Gao, R. Wang, and K. Khalighi, ‘‘Ensemble of machine
achieved accuracies of approximately 87%, 92%, and 95% learning algorithms using the stacked generalization approach to estimate
the Warfarin dose,’’ PLoS ONE, vol. 13, no. 10, pp. 1–12, 2018.
on the first, second, and third datasets, respectively. The high
[20] H. Huang, Y. Cheng, and R. Weibel, ‘‘Transport mode detection based on
accuracy achieved by the proposed method indicates that the mobile phone network data: A systematic review,’’ Transp. Res. C, Emerg.
size of the dataset used to evaluate the proposed method is Technol., vol. 101, pp. 297–312, Apr. 2019.
sufficient. [21] S. Hemminki, P. Nurmi, and S. Tarkoma, ‘‘Accelerometer-based trans-
portation mode detection on smartphones,’’ in Proc. 11th ACM Conf.
Embedded Netw. Sensor Syst., New York, NY, USA, 2013, p. 13.
REFERENCES [22] B. Wang, L. Gao, and Z. Juan, ‘‘Travel mode detection using GPS data and
socioeconomic attributes based on a random forest classifier,’’ IEEE Trans.
[1] B. Schilit, N. Adams, and R. Want, ‘‘Context-aware computing applica- Intell. Transp. Syst., vol. 19, no. 5, pp. 1547–1558, May 2018.
tions,’’ in Proc. 1st Workshop Mobile Comput. Syst. Appl., Piscataway, NJ, [23] J. Urbancic, V. Pejovic, and D. Mladenic, ‘‘Transportation mode detection
USA, Dec. 1994, pp. 85–90. using random forest,’’ in Proc. Inf. Soc., Data Mining Data Warehouses
[2] P. Castrogiovanni, E. Fadda, G. Perboli, and A. Rizzo, ‘‘Smartphone (SiKDD), Ljubljana, Slovenia, 2018, pp. 1–4.
data classification technique for detecting the usage of public or private [24] M. A. Shafique and E. Hato, ‘‘Classification of travel data with multiple
transportation modes,’’ IEEE Access, vol. 8, pp. 58377–58391, 2020. sensor information using random forest,’’ Transp. Res. Procedia, vol. 22,
[3] C. Carpineti, V. Lomonaco, L. Bedogni, M. D. Felice, and L. Bononi, pp. 144–153, Jan. 2017.
‘‘Custom dual transportation mode detection by smartphone devices [25] Z. Xiao, Y. Wang, K. Fu, and F. Wu, ‘‘Identifying different transportation
exploiting sensor diversity,’’ in Proc. IEEE Int. Conf. Pervas. Com- modes from trajectory data using tree-based ensemble classifiers,’’ ISPRS
put. Commun. Workshops (PerCom Workshops), Piscataway, NJ, USA, Int. J. Geo-Inf., vol. 6, no. 2, p. 57, Feb. 2017.
Mar. 2018, pp. 367–372. [26] Y. Zheng, H. Fu, X. Xie, W. Y. Ma, and Q. Li. (2011). Geolife GPS
[4] A. Jahangiri and H. A. Rakha, ‘‘Applying machine learning techniques to Trajectory Dataset-User Guide. Microsoft Research. [Online]. Avail-
transportation mode recognition using mobile phone sensor data,’’ IEEE able: https://www.microsoft.com/en-us/research/publication/geolife-gps-
Trans. Intell. Transp. Syst., vol. 16, no. 5, pp. 2406–2417, Oct. 2015. trajectory-dataset-user-guide/
[5] J. R. Kwapisz, G. M. Weiss, and S. A. Moore, ‘‘Activity recognition using [27] S.-H. Fang, H.-H. Liao, Y.-X. Fei, K.-H. Chen, J.-W. Huang, Y.-D. Lu,
cell phone accelerometers,’’ ACM SIGKDD Explor. Newslett., vol. 12, and Y. Tsao, ‘‘Transportation modes classification using sensors on smart-
no. 2, pp. 74–82, Mar. 2011. phones,’’ Sensors, vol. 16, no. 8, p. 1324, Aug. 2016.
[6] J. Biancat, C. Brighenti, and A. Brighenti, ‘‘Review of transportation mode [28] M. A. Shafique and E. Hato, ‘‘Use of acceleration data for transportation
detection techniques,’’ EAI Endorsed Trans. Ambient Syst., vol. 1, no. 4, mode prediction,’’ Transportation, vol. 42, no. 1, pp. 163–188, Jan. 2015.
p. e7, Oct. 2014. [29] S.-H. Fang, Y.-X. Fei, Z. Xu, and Y. Tsao, ‘‘Learning transportation modes
[7] Y. Qin, H. Luo, F. Zhao, C. Wang, J. Wang, and Y. Zhang, ‘‘Toward from smartphone sensors based on deep neural network,’’ IEEE Sensors J.,
transportation mode recognition using deep convolutional and long vol. 17, no. 18, pp. 6111–6118, Sep. 2017.
short-term memory recurrent neural networks,’’ IEEE Access, vol. 7, [30] X. Liang and G. Wang, ‘‘A convolutional neural network for transportation
pp. 142353–142367, 2019. mode detection based on smartphone platform,’’ in Proc. IEEE 14th Int.
[8] Y. Xiao, D. Low, T. Bandara, P. Pathak, H. Beng Lim, D. Goyal, Conf. Mobile Ad Hoc Sensor Syst. (MASS), Oct. 2017, pp. 338–342.
J. Santos, C. Cottrill, F. Pereira, C. Zegras, and M. Ben-Akiva, ‘‘Trans- [31] L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen, Classification
portation activity analysis using smartphones,’’ in Proc. IEEE Con- Regression Trees. Boca Roton, FL, USA: CRC Press, 1984.
sum. Commun. Netw. Conf. (CCNC), Piscataway, NJ, USA, Jan. 2012, [32] L. Breiman, ‘‘Random forests,’’ Mach. Learn., vol. 45, no. 1, pp. 5–32,
pp. 60–61. Oct. 2001.
[9] L. Stenneth, O. Wolfson, P. S. Yu, and B. Xu, ‘‘Transportation mode [33] J. H. Friedman, ‘‘Stochastic gradient boosting,’’ Comput. Statist. Data
detection using mobile phones and GIS information,’’ in Proc. 19th ACM Anal., vol. 38, no. 4, pp. 367–378, Feb. 2002.
SIGSPATIAL Int. Conf. Adv. Geographic Inf. Syst., New York, NY, USA, [34] T. Cover and P. Hart, ‘‘Nearest neighbor pattern classification,’’ IEEE
2011, pp. 54–63. Trans. Inf. Theory, vol. IT-13, no. 1, pp. 21–27, Jan. 1967.

145562 VOLUME 8, 2020

B. Alotaibi: TMD by Embedded Sensors Based on Ensemble Learning

[35] T. K. Ho, ‘‘The random subspace method for constructing decision [49] B. Nham, K. Siangliulue, and S. Yeung, ‘‘Predicting mode of transport
forests,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 8, from iPhone accelerometer data,’’ Mach. Learn. Final Projects, Stanford
pp. 832–844, Aug. 1998. Univ. Class Project, Stanford, CA, USA, Tech. Rep. CS 229, 2008.
[36] S. C. Kleene, ‘‘Representation of events in nerve nets and finite [50] P. Nitsche, P. Widhalm, S. Breuss, and P. Maurer, ‘‘A strategy on how to uti-
automata,’’ RAND Project Air Force, Santa Monica, CA, USA, lize smartphones for automatically reconstructing trips in travel surveys,’’
Tech. Rep. RAND-RM-704, 1951. Procedia-Social Behav. Sci., vol. 48, pp. 1033–1046, 2012.
[37] Y. Freund and R. E. Schapire, ‘‘A desicion-theoretic generalization of on- [51] O. Burkhard, H. Becker, R. Weibel, and K. W. Axhausen, ‘‘On the require-
line learning and an application to boosting,’’ in Proc. Eur. Conf. Comput. ments on spatial accuracy and sampling rate for transport mode detection in
Learn. Theory. Berlin, Germany: Springer, 1995, pp. 23–37. view of a shift to passive signalling data,’’ Transp. Res. C, Emerg. Technol.,
[38] B. Boehmke and B. M. Greenwell, ‘‘Gradient Boosting,’’ in Hands-On vol. 114, pp. 99–117, May 2020.
Machine Learning With R. Boca Raton, FL, USA: CRC Press, 2019, [52] K. Ellis, S. Godbole, S. Marshall, G. Lanckriet, J. Staudenmayer, and
pp. 221–223. J. Kerr, ‘‘Identifying active travel behaviors in challenging environments
[39] R. Polikar, ‘‘Ensemble learning,’’ in Ensemble Machine Learning. Berlin,
using GPS, accelerometers, and machine learning algorithms,’’ Frontiers
Germany: Springer, 2012, pp. 1–34.
Public Health, vol. 2, p. 36, Apr. 2014.
[40] S. Raschka, ‘‘Stat 479: Machine learning lecture notes,’’ Dept. Statist.,
Algorithm Comparison, Univ. Wisconsin–Madison, Madison, WI, USA, [53] J. A. Carlson, M. M. Jankowska, K. Meseck, S. Godbole, L. Natarajan,
Tech. Rep. STAT479 FS19. L11, 2018. F. Raab, B. Demchak, K. Patrick, and J. Kerr, ‘‘Validity of PALMS GPS
[41] Z. H. Zhou, ‘‘Ensemble learning,’’ Encyclopedia Biometrics, vol. 1, scoring of active and passive travel compared to SenseCam,’’ Med. Sci.
pp. 270–273, Jun. 2009. Sports Exerc., vol. 47, no. 3, p. 662, 2015.
[42] S. Haykin, ‘‘Multilayer perceptron,’’ in Neural Networks: A Compre-
hensive Foundation. Upper Saddle River, NJ, USA: Prentice-Hall, 1994,
pp. 178–191.
[43] M. D. Felice, L. Bononi, L. Bedogni, and V. Lomonaco. (2017). TMD
Dataset. University of Bologna. [Online]. Available: http://cs.unibo.
it/projects/us-tm2017/index.html
[44] H. Gjoreski, M. Ciliberto, L. Wang, F. J. Ordonez Morales, S. Mekki,
S. Valentin, and D. Roggen, ‘‘The university of Sussex-Huawei locomotion
and transportation dataset for multimodal analytics with mobile devices,’’
IEEE Access, vol. 6, pp. 42592–42604, 2018.
[45] H. Xia, Y. Qiao, J. Jian, and Y. Chang, ‘‘Using smart phone sensors to BANDAR ALOTAIBI (Member, IEEE) received
detect transportation modes,’’ Sensors, vol. 14, no. 11, pp. 20843–20865, the B.Sc. degree (Hons.) in computer science-
Nov. 2014. information security and assurance emphasis from
[46] D. M. Karantonis, M. R. Narayanan, M. Mathie, N. H. Lovell, and The University of Findlay, USA, the M.Sc. degree
B. G. Celler, ‘‘Implementation of a real-time human movement classifier in information security and assurance from Robert
using a triaxial accelerometer for ambulatory monitoring,’’ IEEE Trans.
Morris University, USA, and the Ph.D. degree
Inf. Technol. Biomed., vol. 10, no. 1, pp. 156–167, Jan. 2006.
[47] K. Li, M. Lu, F. Lu, Q. Lv, L. Shang, and D. Maksimovic, ‘‘Personalized in computer science and engineering from the
driving behavior monitoring and analysis for emerging hybrid vehicles,’’ University of Bridgeport, USA. He is currently
in Proc. Int. Conf. Pervas. Comput. Berlin, Germany: Springer, 2012, an Assistant Professor with the Department of
pp. 1–19. Information Technology, University of Tabuk. His
[48] D. Shin, D. Aliaga, B. Tunçer, S. M. Arisona, S. Kim, D. Zünd, research interests include computer vision, network security, mobile com-
and G. Schmitt, ‘‘Urban sensing: Using smartphones for transportation munications, computer forensics, wireless sensor networks, and quantum
mode classification,’’ Comput., Environ. Urban Syst., vol. 53, pp. 76–86, computing.
Sep. 2015.

VOLUME 8, 2020 145563