Novel Human Activity Recognition and Recommendation Models
for Maintaining Good Health of Mobile Users
XINYI ZENG1, MENGHUA HUANG1, HAIYANG ZHANG2, ZHANLIN JI3,*, IVAN GANCHEV4,*
1College of Artificial Intelligence,
North China University of Science and Technology,
Tangshan,
CHINA
2Department of Computing,
Xi’an Jiaotong–Liverpool University,
Suzhou,
CHINA
3College of Artificial Intelligence,
North China University of Science and Technology,
Tangshan,
CHINA
also with
Telecommunications Research Center (TRC),
University of Limerick, Limerick,
IRELAND
4University of Plovdiv “Paisii Hilendarski”,
Plovdiv,
BULGARIA
also with
Institute of Mathematics and Informatics,
Bulgarian Academy of Sciences (IMI–BAS),
Sofia,
BULGARIA
also with
Telecommunications Research Center (TRC),
University of Limerick, Limerick,
IRELAND
*Corresponding Authors
Abstract: - With the continuous improvement of the living standard, people have changed their concept from
disease treatment to health management. However, most of the current health management software makes
recommendations based on users’ static information, with low updating frequency. The effect of targeted
suggestions becomes weak with time, and it is hard for the recommendation effect to be satisfactory. Based on
the use of smartphones for recognizing human activities in real-time, firstly, a novel 'CNN+GRU' model is
proposed in this paper, utilizing both convolutional neural networks (CNNs) and gated recurrent units (GRUs).
'CNN+GRU' can improve the recognition speed and extract the features in sensor data more accurately by
achieving in the conducted experiments an average accuracy of 91.27%, thus outperforming other models
compared. Secondly, another model, named SimilRec, is proposed for physical activity recommendation to
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
33
Volume 21, 2024
users based on their health profile, the similarities between their current physical activity sequence, and the
historical physical activity sequence of other (similar) users.
Key-Words: feature extraction; convolutional neural network (CNN); gated recurrent unit (GRU); human
activity recognition (HAR); physical activity recommendation, recommendation system.
Received: July 15, 2022. Revised: October 18, 2023. Accepted: November 18, 2023. Published: January 23, 2024.
1 Introduction
Smartphones and smart wearable devices have
become very popular recently. In addition, lots of
personal health management software has been
developed for such devices. However, the overall
health-related recommendations and suggestions
made are still not well personalized to individual
users. A vast majority of wearable devices worn by
users can only be used to record data that are easy to
obtain, such as the step count or heart rate. Based on
the average daily target of 10,000 steps made, most
commercial software can only calculate the number
of calories burnt by the user in the performed
physical exercises and evaluate his/her health
condition but cannot recognize the actual human
activities and evaluate the real exercise volume of
users. The inability of such software to provide
personal health-related recommendations, that best
suit the user, could badly affect his/her physical
condition and/or behavioral habits.
Human activity recognition (HAR) can be
considered a typical pattern recognition task.
Decision trees, support vector machines (SVMs),
adaptive boosting (AdaBoost), [1], and hidden
Markov models are mainly used for modeling in the
conventional pattern recognition methods, which
have shown good progress and achieved satisfactory
results, [2]. However, these methods are constrained
by human domain knowledge in most daily activity
recognition tasks, [3]. Additionally, these methods
can only be used for learning shallow features,
which may lead to a decline in the HAR
performance.
Deep learning has shown excellent abilities in
various fields in recent years. Different from
conventional pattern recognition methods, it can
greatly reduce the workload of feature design and
help learn more advanced and meaningful features
by training end-to-end neural networks. In addition,
a deep network structure is more suitable for
unsupervised learning. This makes it quite suitable
for HAR. Basically, for this task, the data from
multiple sensors are inputted into a convolutional
neural network (CNN) or recurrent neural network
(RNN) model to obtain time series data and capture
the features therein to recognize human activities.
As illustrated in Figure 1, for the HAR task, sensors
embedded in smartphones and smart wearable
devices (e.g., acceleration sensors, gyroscope
sensors, etc.) can be greatly utilized, [4], to capture
the movement specifics of the corresponding users
and then extract the time series features to recognize
the human activities performed. In the past, these
features were extracted manually. The common
features include statistical features, such as
maximum, mean, and minimum values and
variance, and frequency domain features, such as
Fourier transforms, [5]. Nevertheless, the manually
extracted features are highly dependent on the data
sets used and very poor in generalization, which
necessitates a second manual extraction of features
upon replacement of data sets, which is both time-
and effort-consuming. Additionally, the manual
extraction of features is limited to the cognition of
human experts performing this task. Generally,
experts can extract shallow features but cannot
obtain deep features in most cases. Thus, automatic
extraction of features, e.g., using deep learning, is
mainly used today, [6]. In line with this trend, this
paper considers the use (single or combined) of
artificial neural networks for HAR to find the best
solution by utilizing the data supplied by sensors
embedded in users’ smartphones and/or smart
wearable devices.
Sensor signals Feature extraction
Maximum
value
Mean value
Minimum
value
Variance
Training a model
Decision
tree
AdaBoost
SVM
ANN
Human activity
recognition
(HAR)
... ...
Fig. 1: An illustration of HAR steps
In their normal daily life, people usually follow a
fixed lifestyle by performing various physical
activities in the daytime and having a rest at night.
Important here is the ability to mine users’ life
patterns based on their daily life records, then infer
their subsequent activities based on effective
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
34
Volume 21, 2024
personal historical activity sequences, and
recommend the most appropriate physical activities
to users, so that they can maintain good living habits
and keep their bodies in a good health condition,
[7].
It is possible to recognize basic human activities,
such as walking, running, going upstairs or
downstairs, sitting, standing, etc., by utilizing
models that can recognize human activities based on
mobile phones’ sensor data, [8]. But in daily health
events, the aim is to analyze users' health condition,
based on their eating, sleeping, and other similar
activities, with high-level semantics. To this end,
further analysis is required, based on experiments
with data sets containing the daily activity data of
many users, to analyze the semantic similarities of
the users' activities, [9]. Then, the distance of users'
historical activity sequence can be judged, and the
required subsequent physical activities can be
recommended, as to provide the users with proper
suggestions for maintaining a good health condition.
The activity log data of a user can be easily
recorded if s/he uses a smartphone or a smart
wearable device. Then, when the user is performing
a specific activity, another related activity can be
recommended to be performed next, by analyzing
the user’s historical activity data. The daily activity
logs of some users may be similar, but the duration
of executing each physical activity and the sequence
of activities may be quite different. In addition,
different activities may be executed under special
circumstances. Thus, it is crucial to find possible
influencing factors for executing such activities. In a
daily life log, each activity can be recorded along
with time stamps and other contextual information.
Thus, a life log may be considered as a series of
activity sequences with different features. Each
activity may occur several times within a certain
scope. The entire activity sequence, sorted by the
time of execution, can be expressed as:
where denotes a sequenced set of a series of
activities executed within a
certain period. The different features of the -th
activity , e.g., its start time, duration,
place/location of execution, etc., can be expressed
as:
where denotes the j-th feature of
activity. The feature factors of the sequence need
to be paid attention to when recommending
subsequent activities according to historical activity
records of users. For this, it is necessary to calculate
the similarities between the current activity
sequence and the historical activity sequences
contained in the records. To this end, a similarity
recommendation model, called SimilRec, is
proposed and described further in Subsection 4.2 of
this paper.
The main contributions of the paper could be
summarized as follows:
1) For HAR, a novel 'CNN+GRU' model is
proposed, based on a combined use of CNNs and
Gated Recurrent Units (GRUs). The proposed
model works with human activity data collected by
sensors, embedded in the users’ smartphones or
smart wearable devices. For better performing the
task, the proposed 'CNN+GRU' model is optimized
in terms of the number of utilized convolution
filters, the number of convolution kernels, and the
loss function, in order to find a good compromise
between stability and accuracy achieved. Results,
obtained by experiments conducted on a public data
set, demonstrate that the proposed 'CNN+GRU'
model outperforms other similar models used for
HAR, in terms of the average accuracy achieved.
2) For physical activity recommendation, a novel
SimilRec model is elaborated, which recommends
physical activity to a target user based on his/her
health-related profile and the discovered similarities
between the current physical activity sequence
performed by that user and the historical physical
activity sequence of other (similar) users.
2 Background
2.1 Artificial Neural Networks (ANNs)
ANN is a powerful computing system, consisting of
many artificial neurons connected in a network,
simulating this way various neurons in the human
brain interconnected to perform diverse basic
functions. Overall, each ANN can be divided into an
input layer, a hidden layer, and an output layer. The
hidden layer may have multiple sublayers, which
convert an ANN into a deep neural network (DNN).
Figure 2 shows a fully connected DNN with two
hidden layers, where circles represent the neurons
(a.k.a. activation functions in mathematics) through
which information is transmitted to the next layer.
Introducing more hidden layers into the network
allows it to enhance its capability in performing
different tasks. However, having lots of hidden
layers does not always work well, because this not
only increases the amount of computation but also
leads to gradient explosion or gradient
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
35
Volume 21, 2024
disappearance, resulting in worse network
performance compared to using only fewer layers,
[10].
Input Layer Hidden Layer 1
Output Layer
Hidden Layer 2
Fig. 2: A fully connected DNN with two hidden
layers
The two main ANN types used in the models,
proposed in this paper, are CNNs and RNNs,
described in the following subsections.
2.1.1 Convolutional Neural Networks (CNNs)
CNNs were developed in response to image
classification problems. A CNN model can process
multi-dimensional sensor data series, extract
features from such series, and map internal features
to different activity types, [11]. The advantage of
using CNNs in classification tasks relates to the
direct extraction of features from the original series.
A typical one-dimensional (1D) CNN model
with a sequential structure is depicted in Figure 3,
[12]. The convolution layer is essentially a feature
extraction layer. A hyperparameter is set to
specify how many feature extractors (i.e., filters that
process the input data in parallel) are used. The
flatten layer is used as a transition from the
convolution layer to the fully connected layer to
flatten multi-dimensional data into 1D data, [13].
The pooling layer is used to reduce the number of
feature samples to a quarter of the original number,
highlighting the most obvious features. The pooling
layer performs dimensionality reduction operations
on the features of the filter to form the final features.
The dropout layer is used for data normalization.
Due to the high learning speed of CNNs, the
dropout layer is needed to help slow down the
learning, prevent the model from overfitting, and
improve the generalization ability of the model.
Generally, a fully connected layer is used after that
to complete the classification process. For the
multitype task of HAR, the Adam optimization
algorithm is usually used to optimize the network,
and the loss function adopts the categorical cross-
entropy loss.
Input layer Convolution
layer Pooling
layer
Filters
DropOut
Fully connected layer
Output
layer
Flatten
layer
…
Fig. 3: A typical 1D CNN structure
2.1.2 Recurrent Neural Networks (RNNs)
RNNs are a special type of neural network, designed
to process time series data. RNNs are widely used
for speech recognition and machine translation,
[14]. The continuous inputs of RNN are correlated
with each other. However, there are many
cumulative products, which may easily cause the
problem of vanishing or exploding gradients. In
addition, RNNs have higher requirements for the
hardware used, and their real-time performance for
pattern recognition is low. A typical RNN structure
is shown in Figure 4. Note that removing the
layer converts the RNN into a fully connected
neural network.
In Figure 4, represents the vector in the input
layer, the vector of the intermediate
hidden layer, and represents the vector of the
output layer. S can be calculated as follows:
where denotes the weight matrix from the input
layer to the hidden layer, denotes the weight
matrix from the hidden layer to the output layer in
which , denotes the weight of the
hidden layer, and denotes the vector of the
previous hidden layer. This way, the previous data
are used in each cycle. However, if there are too
many cycles, will be multiplied several times.
Then, this may lead to the problem of exploding or
vanishing gradients, which can be solved by a new
type of network, called long short-term memory
(LSTM), which is formed by expanding the RNN
structure, as depicted in Figure 5.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
36
Volume 21, 2024
X
S
O
Input layer
Hidden layer
Output layer
W
Fig. 4: A typical RNN structure
X
S
O
Input layer
Hidden layer
Output layer
WW
Xt-1
U
V
St-1
U U
V V
StSt+1
XtXt+1
Ot-1 OtOt+1
Fig. 5: The expansion of a RNN into a LSTM
2.1.3 LSTMs
With the improvement of technology, LSTMs have
gradually evolved from RNNs, [15]. LSTMs are
mainly used for natural language processing (NLP)
and time series data processing. A LSTM model can
not only avoid the problems of vanishing and
exploding gradients, but also can keep the features
in the time series sequence from getting lost, thanks
to the three gates utilized, [16], as shown in Figure
6. The input gate is used to form the current input,
using a sigmoid function, and add it to the value of
the previous hidden state to maintain the long-term
features of the time series sequence unchanged. The
forget gate determines which information should be
lost or retained. The output gate determines what the
next hidden state is.
σ σ tanh σ
x(t)
h(t-1)
c(t-1)
h(t)
c(t)
h(t)
× +
×
tanh
×
Fig. 6: A typical LSTM structure
2.1.4 Gated Recurrent Units (GRUs)
As a variant of the LSTM, the gated recurrent unit
(GRU) combines the forget gate and the input gate
into a single update gate, which allows to reduce the
parameter calculation workload and shorten the
model’s training time, [17]. For both GRU and
LSTM, gates can be used to retain important
features. The selection of GRU or LSTM for use is
generally determined by the specific task.
Relatively, as shown in Figure 7, GRU has a simpler
structure than LSTM, which means a smaller
calculation workload. As a result, GRU maintains a
fast calculation speed when a large amount of input
data is used.
σ σ tanh
x(t)
h(t-1) h(t)
× +
×
1-
×
h(t)
r(t) z(t) ht’
Fig. 7: A typical GRU structure.
In the proposed 'CNN+GRU' model, presented in
Subsection 4.1, GRUs are used as units for
extracting time series features from the collected
sensor data. The experimental results, presented in
Subsection 5.2, show that the combined use of
CNN(s) and GRU(s), as utilized in the proposed
model, allows to achieve better results in the HAR
task, when compared to the combined use of
CNN(s) and LSTM(s), or the single use of LSTM or
GRU.
2.2 User Profiles
Health-status-related user profiles contain the users’
basic health attributes and health preferences. The
initial user profiles are usually created by utilizing
the information entered by the respective users at
the time of their registration in health-related
information systems. Subsequently, these profiles
are continuously updated in the process of using
such systems by the users, either explicitly by
utilizing the feedback (preferences, reviews,
comments, tags, etc.) provided directly by the users
(through various means) or implicitly by assessing
the user behavior in using these systems. The
updated user profiles are in turn utilized for the
provision of more accurate recommendations of
daily activities and/or physical exercises to the
respective users. The process of creating and
updating a user profile to recommend physical
activities is shown in Figure 8.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
37
Volume 21, 2024
User
feedback User
profile
User
information
Recommended
physical
activities
User
Recommendation
model
update
enter
create
provide update
supply
list
prepare
list
Fig. 8: The process of creating and updating a user's
profile to recommend physical activities.
A user profile is initially created by inputting the
user's personal information, such as gender, age,
height, weight, athletic goal, and level, etc.,
considered as different features of that particular
user, i.e.:
where denotes the -th feature of the -th user .
One-hot coding is usually carried out in the selected
areas for features, to reduce the data dimension and
lessen the computation difficulty.
3 Related Work
3.1 Human Activity Recognition (HAR)
Due to the rapid development of the Internet of
Things (IoT) and ubiquitous computing, human
activity recognition based on various sensors is
becoming more and more popular. Many HAR
examples can be found nowadays. For instance,
bracelets with an integrated heart rate sensor, an
acceleration sensor, and other sensors, can capture
rich human motion data. Sensors in smartphones are
even more abundant. Mobile crowd sensing and
computing methods, utilizing the massive number of
mobile sensing devices and personal communication
devices (smartphones, tablets, smartwatches, etc.),
are usually used for collecting the needed data and
reducing the cost of data acquisition. However, due
to the uneven data quality and reliability of different
processing methods used, effective data analysis and
mining should be carried out subsequently, e.g., by
employing deep learning techniques. There are
many methods to recognize human activities
through mobile phones’ sensors, and multiple
corresponding applications exist, e.g., for fall
detection, step count statistics, and so on.
3.2 Physical Activity Recommendation
In the case of everyday health events, the goal is to
analyze the user's health status based on daily life
events, such as eating, sleeping, etc., with high-level
semantics. At present, recommendation systems can
be roughly divided into the following categories:
1) Recommendation systems based on user
behavior: These systems employ traditional
collaborative filtering recommendation algorithms
and matrix decomposition algorithms. Through
regular user behavior analysis and model training,
user characteristic information and model
parameters are updated accordingly;
2) Tag-based recommendation systems: These
systems do not need complex algorithms. For
instance, in an e-commerce recommendation
system, users can recommend resources they are
interested in according to their associated tags; in
social networks, users can find friends with the
same hobbies through tags, etc.;
3) Recommendation systems utilizing deep
learning models: Compared with traditional
recommendation models, deep learning models
generally have stronger feature expression and
generalization ability, can effectively capture
nonlinear and unusual relationships between users
and resources, and support more complex
abstractions as higher-level data representation. In
addition, they can learn complex relationships in the
available data, contextual text, and visual
information.
3.3 Recommendation Approaches
The goal of computer algorithms, used by the
respective recommendation systems, is to provide
users with accurate recommendations as fast as
possible, [18]. Different recommendation approaches
exist, each with its pros and cons, due to different
recommendation tasks and different types of data
sources utilized. Generally, the recommendation
approaches could be divided into two main groups –
content-based filtering (CBF) and collaborative
filtering (CF).
CBF, [19], depends on the resource portrait and
user behavior. It can search for similar resources
(e.g., items, services, physical activities, exercises,
etc.) under the portrait information of resources in the
user history and recommend them to the user. CBF is
widely used in industry due to its simplicity and
efficiency.
CF, however, is typically favored over CBF due to
its overall better performance in predicting common
behavior patterns and its ability to address data
aspects that are usually difficult to profile using CBF,
[20]. CF only requires user–resource interactions to
make recommendations, meaning that it is easier to
adapt it to real-world scenarios than CBF, [21]. As a
result, CF has been more successful and more widely
used than CBF, as it only relies on the past user’s
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
38
Volume 21, 2024
behavior (e.g., previous user’s transactions, reviews,
ratings, tags, etc.) without requiring specific domain
information.
A widely accepted taxonomy, [22], divides CF
into two categories:
(1) Memory-based CF, in which
recommendations are based on the assumption that
users who share common interests have similar tastes,
or resources with similar features have similar rating
patterns;
(2) Model-based CF, in which various machine
learning or data mining techniques are utilized to
discover complex patterns in the user history data to
make recommendations based on these, [18], [21].
Generally, the memory-based CF is simpler and
easy to implement, while at the same time, it can
obtain reasonably accurate results, [18]. Two of the
most widely used methods in this category are based
on the k-nearest neighbor (KNN) heuristic, [23],
divided into: (i) user-based KNN, [24], that predicts
the rating of resource i by user u by using the existing
ratings given to i by the set of users that are most
similar to u; and (ii) resource-based KNN, [25], [26]
that predicts the rating of resource i by user u using
the existing ratings given by u to the set of resources
that are most similar to resource i. Both methods use
the following two steps to make predictions:
(1) Finding the k most similar neighbors to the
target user/resource: The most important part here is
to compute the similarity between users/resources.
The two most popular choices for similarity metrics
are: (i) the Pearson’s correlation coefficient, [23],
which measures the extent to which two vectors are
linearly related to each other; and (ii) the cosine
similarity, [27], which measures the similarity
between two vectors by computing the cosine of the
angle between them, [28];
(2) Aggregating the neighbors to generate the
predictive score, [18], [23]: The predicted rating is
calculated based on the ratings of the k-nearest
neighbors selected in the first step, e.g., as the
weighted sum of the ratings of the same resource,
provided by the neighbors of the target user, or as the
average sum of other ratings.
Due to its simplicity and flexibility, [23], the
nearest-neighbor-based CF has been extensively
studied, including different similarity measures, [29]
[30], [31], alternative strategies to select the
neighbors, etc. Two drawbacks of the memory-based
CF are: (i) the low efficiency since the computation
of the similarity between users/resources is expensive
(quadratic time complexity) as all users/resources
need to be examined to make a single prediction, and
(ii) the performance of recommendation heavily
depends on the similarity measure, [32].
On the other side, the model-based CF can tackle
the data sparsity and scalability issues that the
memory-based CF struggles to cope with. In addition,
the model-based CF can achieve better
recommendation performance and coverage than the
memory-based CF, because it trains a model based on
global rating data, while the memory-based CF only
focuses on the local rating information, [33]. Various
machine learning and data mining algorithms have
been elaborated by different researchers in the past
for making recommendations, such as Restricted
Boltzmann Machines, [34], regression-based models,
[35] and latent factor models (mostly based on matrix
factorization, [20], e.g., SVD [36], SVD++ [37]), etc.
The elaborated SimilRec model, presented in
Subsection 4.2, utilizes model-based CF techniques.
More specifically, it is based on the word2vec model,
[38] and the Continuous Bag-of-Words (CBOW)
model, [39].
4 Proposed Models
4.1 'CNN+GRU' (for Human Activity
Recognition)
For HAR, a novel 'CNN+GRU' model is proposed
here, based on a combined use of CNN and GRU.
The elaborated 'CNN+GRU' model consists of three
parts (Figure 9): the first part is used to extract
features using two convolution layers; the second
part is used to obtain the time series relationship
existing in the collected sensor data through two
GRU layers; and the third part is used to expand the
data generated by GRU using a fully connected
layer, then input all data into a SoftMax function,
and finally get the classification result of human
activities.
One important hyperparameter in the CNN part
of the proposed model is the number of filters used,
initially being set to 8. However, we experimented
also with other values, such as 16, 32, 64, 128, and
256, to find the optimal value of this
hyperparameter for the proposed model. The
obtained results are shown in Figure 10. The
presented box plot diagram shows that the highest
median classification accuracy is achieved with 128
filters used; however, the stability then is not great.
Thus, as a good compromise between stability and
accuracy, the default value of convolution filters is
set to 64 in the proposed 'CNN+GRU' model.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
39
Volume 21, 2024
batch×128×9
batch×126×32 batch×63×32 batch×63×64
.
.
.
.
.
batch×4032 batch×6
Convolution layer + ReLU function
Pooling layer
Bonding layer
GRU
Fully connected layer
SoftMax
batch×63×64
Input layer
Fig. 9: The structure of the proposed 'CNN+GRU'
model for human activity recognition
Fig. 10: The box plot diagram of accuracy vs. the
number of convolution filters used by the proposed
'CNN+GRU' model
Another important hyperparameter of the
proposed model is the number of convolution
kernels, used to control the time step for
computation each time the input sequence is read
and then mapped to filters through convolution. A
bigger number of kernels can embrace a wider data
range for processing, thus achieving higher
accuracy. However, the increase in the number of
kernels leads to instability. To obtain the optimal
value of this hyperparameter, experiments were
conducted with different numbers of kernels,
namely 2, 3, 5, 7, and 11. The obtained results are
shown in Figure 11. According to the presented box
plot diagram, the highest accuracy is achieved with
11 kernels; however, the stability then is not good.
Thus, as a compromise between stability and
accuracy, the default value of convolution kernels is
set to 5 in the proposed 'CNN+GRU' model.
Fig. 11: The box plot diagram of accuracy vs. the
number of convolution kernels used by the proposed
'CNN+GRU' model
The proposed model works with human activity
data collected by sensors, embedded in the
smartphones or smart wearable devices of users.
These data are first converted into proper time
series. By inputting types of sensor data, the
length of each sensor data sequence becomes . So,
the input to the convolution layers is:
(5)
where denotes the time series
corresponding to the types of sensor data at time
. 1D convolution layers, each adopting a ReLU
as the activation function, are used. This allows the
model to learn more complex features, which can be
subdivided further so that the combined data may
have more complex feature structures. The result of
the convolution is:
where denotes the weight of each filter and
denotes the relevant offset value. After the
convolution, the data are processed by the pooling
layer. The pooling process helps process data with
relatively large fluctuations, which allows the model
to more accurately retain signal fluctuations while
extracting features. Then, the pooled data structures
are merged and inputted into the two GRU layers,
which are used to learn the correlations existing
between the sensor sequence data. Each GRU layer
is composed of basic GRUs. The output of the-
th GRU at time is:
where the output of the convolution
layer at time , and denotes the output of
the-th GRU at time .
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
40
Volume 21, 2024
In order to prevent the model from overfitting
after passing the two-layer GRU, dropout is used to
suspend the operation of some units, thus making
the model more generalized. Using two GRU layers
by the proposed 'CNN+GRU' model is sufficient as
having more GRU layers would lead to an
exponential growth of memory overhead and time
overhead. Moreover, the vanishing gradient problem
and the dilemma of local optimality may also occur
between different GRU layers, [40].
After passing the second GRU layer, the features
of the time series are further enhanced, and more
accurate features can be extracted. Then, the GRU
output values are expanded and inputted into the
fully connected layer. The learned features are
mapped to the labeled space to form a 1D array.
Finally, the final classification result is obtained by
applying a SoftMax function to the output as
follows:
where denotes the output values of the fully
connected layer.
The proposed 'CNN+GRU' model is optimized
by using the cross-entropy loss function:
wherethe true value in the sample class
.
4.2 SimilRec (for Physical Activity
Recommendation)
The SimilRec model, proposed in this paper for
physical activity recommendation, belongs to the
model-based CF category. To recommend a physical
activity to a target user based on his/her health
profile (containing among other things the physical
activity records of that user), the similarities
between the current physical activity sequence of
the user and the historical physical activity sequence
of other (similar) users is calculated first in three
steps:
(1) Calculating the correlation between
(historical) physical activity sequences
(of the target user and all other users).
Each physical activity sequence, containing
different physical activities performed daily by the
corresponding user, can be regarded as a paragraph
of statements composed of a certain number of
words. By utilizing the word2vec model, [38] and
the CBOW model, [39], each physical activity
sequence is transformed into an activity vector.
First, the continuous word bag concept of word2vec
is used to model the physical activity sequences.
Then, the CBOW model is used to predict the central
word according to the context words. The CBOW
model consists of three layers – an input layer, a
hidden layer, and an output layer –, as depicted on
Figure 12.
Input layer
WV×N
Hidden layer Output layer
W
’
N×V
WV×N
WV×N
hyj
x1k
x2k
xCk
w(t)
C×V-dim
N-dim V-dim
Fig. 12: The CBOW model structure
Given the number of contextual words , the
dimension of the word vector space , an input
layer with one-hot codes for contextual words, the
initial input weight matrix , and the output
weight matrix , the output of the hidden
layer can be calculated as:
where denotes the input of each node. Then, the
output vector can be calculated as:
The input weight matrix and the output
weight matrix are repeatedly updated with
the decrease in the subsequent gradient. The word
vector of the one-hot code can be obtained by
multiplying the vector of the code with the input
weight matrix. This way, the activity vectors of all
users can be obtained, based on their physical
activity sequences. Then, the cosine similarity
is used to measure the correlation between
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
41
Volume 21, 2024
the activity vector of the target user and the activity
vector of each other user, as follows:
where and denote the corresponding
components of the activity vectors of the two users.
The larger the cosine similarity obtained, the
stronger the correlation between the physical
activity sequences of the two users (within the
period considered).
(2) Calculating the edit distance between physical
activities (contained in the historical physical
activity sequences of the target user and all other
users).
Each physical activity has features such as name,
start time, duration, place/location where it has been
carried out by the corresponding user, etc. The edit
distance between two physical activities and
can be calculated by counting the minimum
number of steps required to transform into by
changing (editing) the features of (one feature
per step). If the edit distance between two physical
activities is relatively low, the correlation between
these activities is high.
(3) Calculating the Levenshtein distance between
physical activity sequences (of the target user and
all other users).
The Levenshtein distance, [41], is a string metric
for measuring the difference between two strings
(sequences). For instance, the Levenshtein distance
between two words is represented by the minimum
number of single-character edits (insertions,
deletions, substitutions) required to convert one
word into the other. A lower Levenshtein distance
indicates a bigger similarity between two strings.
The Levenshtein distance between the
first i characters in string and the first j characters
in string is defined as:
Based on the Levenshtein distance, similar
physical activity sequences and can be
discovered in the historical records of two users
(i.e., the target user and another user) by counting
the minimum number of operations required to
transform into . In this process: (i) a physical
activity is inserted into with distance ; (ii)
an activity in is substituted with another activity
with distance ; and (iii) an activity is
deleted from with distance . Then, the
activity distance between two physical activity
sequencesand can be expressed as:
ert
where , , and denote how many times a physical
activity was inserted into , deleted from , and
substituted in with another activity, respectively.
After obtaining the activity distance
between the current activity
sequence of the target user (discovered in real
time by a HAR technique) and each sequence
(found in the historical record of some other user),
the score
for an activity (carried
out by that other user), which is a candidate for
recommendation to the target user, is calculated as:
where
denotes the maximum
activity distance among all distances existing
between the current activity sequence (of the
target user) and each other activity sequence
(contained in the historical records of users). This is
then repeated to each other user, who is different to
the target user.
A score list of physical activities, which are
candidates for recommendation to the target user, is
prepared at the end, based on (15). Then, the
average score for all same-name physical activities
is calculated, followed by preparing a ranked list of
all candidate activities, based on their average score.
Finally, the first N activities in this list are
recommended to the target user.
5 Experiments and Results
5.1 Data Set
The data set used in the conducted experiments was
a public data set for human activity recognition
using smartphones, [42], available from the machine
learning repository of the University of California
Irvine (UCI) [43], called here UCI data set for short.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
42
Volume 21, 2024
This data set contains data of 30 adult volunteers
performing activities of daily living while carrying a
waist-mounted smartphone with embedded sensors.
Through six types of activities (i.e., walking
horizontally, walking upstairs, walking downstairs,
sitting, standing, and lying down), the 3-axial linear
acceleration and 3-axial angular velocity were
captured by the smartphone’s accelerometer and
gyroscope of each participant at a constant rate of
50 Hz. The sensor signals were pre-processed using
noise filters and then sampled in 128 time-step
sliding windows of 2.56 sec with 50% overlap. The
sensor acceleration signal was separated using a
Butterworth low-pass filter into body acceleration
and gravity components. The data set was randomly
partitioned into two sets – a training set, containing
70% of the volunteers’ data, and a test set,
containing the rest of the data.
Figure 13, Figure 14, and Figure 15 depict
sample time series of the X-axis linear acceleration,
extracted from this data set, corresponding to three
recorded human activities, respectively. The
presented data have obvious differences of course.
In the case of walking activity (Figure 13), the data
fluctuates violently with quite a large amplitude; the
waveform, however, is quite regular. In the case of
standing activity (Figure 14), there are only
occasional fluctuations in data; the overall curve
hardly moves, indicating that the participant is in a
static state. In the case of the walking upstairs
activity (Figure 15), although the data fluctuations
are similar to that in the walking activity case, the
overall regularity is not the same, indicating that this
is a different activity indeed.
Fig. 13: Sample time series of the X-axis linear
acceleration corresponding to walking activity
(based on the public UCI data set)
Fig. 14: Sample time series of the X-axis linear
acceleration corresponding to standing activity
(based on the public UCI data set)
Fig. 15: Sample time series of the X-axis linear
acceleration corresponding to walking upstairs
activity (based on the public UCI data set)
5.2 Results
The public UCI data set was used to conduct
performance comparison experiments with four
models, shown in Table 1. The models were
compared in terms of the average accuracy achieved
by each of them in the HAR task. The obtained
results (Table 1), demonstrate that the 'CNN+GRU'
model, proposed in this paper, outperforms all other
models, i.e., LSTM, GRU, and 'CNN+LSTM'. In
addition, it is evident from the results that
combining multiple types of neural networks has a
better effect on HAR than using a single neural
network. In addition, the results show that
combining CNN(s) with GRU(s) is better for HAR
than combining CNN(s) with LSTM(s).
Table 1. The HAR performance of compared
models
Model
Average
accuracy (%)
LSTM
88.62
GRU
88.33
‘CNN+LSTM’
89.25
‘CNN+GRU’
(proposed)
91.27
6 Conclusion
This paper has presented a combined use of
convolutional neural networks (CNNs) and gated
recurrent units (GRUs) for building a novel model,
named ‘CNN+GRU’, for human activity
recognition. Multiple CNNs were adopted to extract
sensor data features and capture more detailed
information. Next, GRUs were used to extract the
time series relationship between data features.
Compared with the traditional single, simple neural
networks, the recognition accuracy has been
improved, the values of model parameters have been
reduced, and both training and recognition speeds
have been increased. As a result, the user can upload
the mobile phone’s sensor data to a server, generate
the activity vector, and calculate the correlation
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
43
Volume 21, 2024
between his/her (historical) physical activity
sequences and those of other users by using the
second model, proposed in this paper, called
SimilRec. A score list of physical activities, which
are candidates for recommendation to the target
user, is prepared at the end and the average score for
all same-name physical activities is calculated by
SimilRec, followed by the preparation of a ranked
list of candidate activities, based on their average
score. Finally, the first N activities in this list are
recommended to the target user.
References:
[1] S. Wu and H. Nagahashi, "Parameterized
AdaBoost: Introducing a Parameter to Speed
Up the Training of Real AdaBoost," IEEE
Signal Processing Letters, vol. 21, no. 6, pp.
687-691, 2014, doi:
10.1109/LSP.2014.2313570.
[2] K. Chen, L. Yao, D. Zhang, X. Wang, X.
Chang, and F. Nie, "A Semisupervised
Recurrent Convolutional Attention Model for
Human Activity Recognition," IEEE
Transactions on Neural Networks and
Learning Systems, vol. 31, no. 5, pp. 1747-
1756, 2020, doi:
10.1109/TNNLS.2019.2927224.
[3] S. Angerbauer, A. Palmanshofer, S. Selinger,
and M. Kurz, "Comparing Human Activity
Recognition Models Based on Complexity
and Resource Usage," Applied Sciences, vol.
11, no. 18, 2021, doi: 10.3390/app11188473.
[4] B. Fu, N. Damer, F. Kirchbuchner, and A.
Kuijper, "Sensing Technology for Human
Activity Recognition: A Comprehensive
Survey," IEEE Access, vol. 8, pp. 83791-
83820, 2020, doi:
10.1109/ACCESS.2020.2991891.
[5] H. Ku, P. Zhou, X. Cai, H. Yang, and Y.
Chen, "Person re-identification method based
on CNN and manually-selected feature
fusion," in 2017 13th International
Conference on Natural Computation, Fuzzy
Systems and Knowledge Discovery (ICNC-
FSKD), 29-31 July 2017 2017, pp. 93-96, doi:
10.1109/FSKD.2017.8393401.
[6] C. Ding, Y. Jia, G. Cui, C. Chen, X. Zhong,
and Y. Guo, "Continuous Human Activity
Recognition through Parallelism LSTM with
Multi-Frequency Spectrograms," Remote
Sensing, vol. 13, no. 21, 2021, doi:
10.3390/rs13214264.
[7] C. T. Yen, J. X. Liao, and Y. K. Huang,
"Human Daily Activity Recognition
Performed Using Wearable Inertial Sensors
Combined With Deep Learning Algorithms,"
IEEE Access, vol. 8, pp. 174105-174114,
2020, doi: 10.1109/ACCESS.2020.3025938.
[8] M. Cornacchia, K. Ozcan, Y. Zheng, and S.
Velipasalar, "A Survey on Activity Detection
and Classification Using Wearable Sensors,"
IEEE Sensors Journal, vol. 17, no. 2, pp. 386-
403, 2017, doi: 10.1109/JSEN.2016.2628346.
[9] L. Dhammi and P. Tewari, "Classification of
Human Activities using data captured through
a smartphone using deep learning techniques,"
in 2021 3rd International Conference on
Signal Processing and Communication
(ICPSC), 13-14 May 2021 2021, pp. 689-694,
doi: 10.1109/ICSPC51351.2021.9451772.
[10] K. D. Apostolidis and G. A. Papakostas, "A
Survey on Adversarial Deep Learning
Robustness in Medical Image Analysis,"
Electronics, vol. 10, no. 17, 2021, doi:
10.3390/electronics10172132.
[11] I. H. Hsieh, H.-C. Cheng, H.-H. Ke, H.-C.
Chen, and W.-J. Wang, "A CNN-Based
Wearable Assistive System for Visually
Impaired People Walking Outdoors," Applied
Sciences, vol. 11, no. 21, 2021, doi:
10.3390/app112110026.
[12] H. Cheng, Z. Xie, Y. Shi, and N. Xiong,
"Multi-Step Data Prediction in Wireless
Sensor Networks Based on One-Dimensional
CNN and Bidirectional LSTM," IEEE Access,
vol. 7, pp. 117883-117896, 2019, doi:
10.1109/ACCESS.2019.2937098.
[13] S. Mekruksavanich and A. Jitpattanakul, "A
Multichannel CNN-LSTM Network for Daily
Activity Recognition using Smartwatch
Sensor Data," in 2021 Joint International
Conference on Digital Arts, Media and
Technology with ECTI Northern Section
Conference on Electrical, Electronics,
Computer and Telecommunication
Engineering, 3-6 March 2021 2021, pp. 277-
280, doi:
10.1109/ECTIDAMTNCON51128.2021.9425
769.
[14] S. Zhou and T. Gao, "Brain Activity
Recognition Method Based on Attention-
Based RNN Mode," Applied Sciences, vol. 11,
no. 21, 2021, doi: 10.3390/app112110425.
[15] A. Shrestha, H. Li, J. L. Kernec, and F.
Fioranelli, "Continuous Human Activity
Classification From FMCW Radar With Bi-
LSTM Networks," IEEE Sensors Journal, vol.
20, no. 22, pp. 13607-13619, 2020, doi:
10.1109/JSEN.2020.3006386.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
44
Volume 21, 2024
[16] L. Zhang, C. Xu, Y. Gao, Y. Han, X. Du, and
Z. Tian, "Improved Dota2 lineup
recommendation model based on a
bidirectional LSTM," Tsinghua Science and
Technology, vol. 25, no. 6, pp. 712-720, 2020,
doi: 10.26599/TST.2019.9010065.
[17] T. Ergen and S. S. Kozat, "Unsupervised
Anomaly Detection With LSTM Neural
Networks," IEEE Transactions on Neural
Networks and Learning Systems, vol. 31, no.
8, pp. 3127-3141, 2020, doi:
10.1109/TNNLS.2019.2935975.
[18] M. Jalili, S. Ahmadian, M. Izadi, P. Moradi,
and M. Salehi, "Evaluating Collaborative
Filtering Recommender Algorithms: A
Survey," IEEE Access, vol. 6, pp. 74003-
74024, 2018, doi:
10.1109/ACCESS.2018.2883742.
[19] M. J. Pazzani and D. Billsus, "Content-based
recommendation systems," in The adaptive
web: Springer, 2007, pp. 325-341.
[20] Y. Koren, R. Bell, and C. Volinsky, "Matrix
Factorization Techniques for Recommender
Systems," Computer, vol. 42, no. 8, pp. 30-37,
2009, doi: 10.1109/MC.2009.263.
[21] J. Bobadilla, F. Ortega, A. Hernando, and A.
GutiéRrez, "Recommender systems survey,"
Know.-Based Syst., vol. 46, pp. 109-132,
2013, doi: 10.1016/j.knosys.2013.03.012.
[22] Y. Koren, "Factorization meets the
neighborhood: a multifaceted collaborative
filtering model," presented at the Proceeding
of the 14th ACM SIGKDD international
conference on Knowledge discovery and data
mining - KDD '08, Las Vegas, Nevada, USA,
2008.
[23] X. Ning, C. Desrosiers, and G. Karypis, "A
Comprehensive Survey of Neighborhood-
Based Recommendation Methods," in
Recommender Systems Handbook, F. Ricci, L.
Rokach, and B. Shapira Eds. Boston, MA,
US.: Springer, 2015, ch. Chapter 2, pp. 37-76.
[24] J. L. Herlocker, J. A. Konstan, A. Borchers,
and J. Riedl, "An algorithmic framework for
performing collaborative filtering," presented
at the Proceedings of the 22nd annual
international ACM SIGIR conference on
Research and development in information
retrieval - SIGIR '99, Berkeley, California,
USA, 1999.
[25] B. Sarwar, G. Karypis, J. Konstan, and J.
Reidl, "Item-based collaborative filtering
recommendation algorithms," presented at the
Proceedings of the tenth International
Conference on World Wide Web - WWW '01,
Hong Kong, 2001.
[26] G. Linden, B. Smith, and J. York,
"Amazon.com recommendations: item-to-item
collaborative filtering," IEEE Internet
Computing, vol. 7, no. 1, pp. 76-80, 2003, doi:
10.1109/mic.2003.1167344.
[27] P. Resnick, N. Iacovou, M. Suchak, P.
Bergstrom, and J. Riedl, "GroupLens: an open
architecture for collaborative filtering of
netnews," presented at the Proceedings of the
1994 ACM conference on Computer
supported cooperative work - CSCW '94,
Chapel Hill, North Carolina, USA, 1994.
[28] G. Salton and M. J. McGill, Introduction to
Modern Information Retrieval. McGraw-Hill,
Inc., 1986.
[29] G. Guo, J. Zhang, and N. Yorke-Smith, "A
Novel Evidence-Based Bayesian Similarity
Measure for Recommender Systems," ACM
Transactions on the Web, vol. 10, no. 2, pp. 1-
30, 2016, doi: 10.1145/2856037.
[30] H. J. Ahn, "A new similarity measure for
collaborative filtering to alleviate the new user
cold-starting problem," Inform Sciences, vol.
178, no. 1, pp. 37-51, 2008, doi:
10.1016/j.ins.2007.07.024.
[31] H. Liu, Z. Hu, A. Mian, H. Tian, and X. Zhu,
"A new user similarity model to improve the
accuracy of collaborative filtering," Know.-
Based Syst., vol. 56, no. C, pp. 156-166, 2014.
[32] Y. Shi, M. Larson, and A. Hanjalic,
"Collaborative Filtering beyond the User-Item
Matrix: A Survey of the State of the Art and
Future Challenges," (in English), ACM
Comput. Surv., vol. 47, no. 1, pp. 1-45, 2014,
doi: 10.1145/2556270.
[33] G. Guo, J. Zhang, and D. Thalmann,
"Merging trust in collaborative filtering to
alleviate data sparsity and cold start,"
Knowledge-Based Systems, vol. 57, pp. 57-68,
2014, doi: 10.1016/j.knosys.2013.12.007.
[34] R. Salakhutdinov, A. Mnih, and G. Hinton,
"Restricted Boltzmann machines for
collaborative filtering," presented at the
Proceedings of the 24th international
conference on Machine learning - ICML '07,
Corvalis, Oregon, USA, 2007.
[35] S. Vucetic and Z. Obradovic, "Collaborative
Filtering Using a Regression-Based
Approach," Knowledge and Information
Systems, vol. 7, no. 1, pp. 1-22, 2005, doi:
10.1007/s10115-003-0123-8.
[36] B. Sarwar, G. Karypis, J. Konstan, and J.
Riedl, "Application of dimensionality
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
45
Volume 21, 2024
reduction in recommender system-a case
study," Minnesota Univ Minneapolis Dept of
Computer Science, 2000.
[37] Y. Koren, "Factorization meets the
neighborhood: a multifaceted collaborative
filtering model," in Proceedings of the 14th
ACM SIGKDD international conference on
Knowledge discovery and data mining, 2008,
pp. 426-434.
[38] Y. Chen, S. Huang, H. Lee, Y. Wang, and C.
Shen, "Audio Word2vec: Sequence-to-
Sequence Autoencoding for Unsupervised
Learning of Audio Segmentation and
Representation," IEEE/ACM Transactions on
Audio, Speech, and Language Processing,
vol. 27, no. 9, pp. 1481-1493, 2019, doi:
10.1109/TASLP.2019.2922832.
[39] T. Mikolov, K. Chen, G. Corrado, and J.
Dean, "Efficient Estimation of Word
Representations in Vector Space," 2013.
[40] Y. Deng, L. Wang, H. Jia, X. Tong, and F. Li,
"A Sequence-to-Sequence Deep Learning
Architecture Based on Bidirectional GRU for
Type Recognition and Time Location of
Combined Power Quality Disturbance," IEEE
Transactions on Industrial Informatics, vol.
15, no. 8, pp. 4481-4493, 2019, doi:
10.1109/TII.2019.2895054.
[41] V. I. Levenshtein, "Binary Codes Capable of
Correcting Deletions, Insertions, and
Reversals," Soviet Physics Doklady, vol. 10,
p. 707, 1966.
[42] D. Anguita, A. Ghio, L. Oneto, X. Parra, and
J. L. Reyes-Ortiz, "A Public Domain Dataset
for Human Activity Recognition using
Smartphones," in ESANN, 2013.
[43] J. Reyes-Ortiz, D. Anguita, A. Ghio, L.
Oneto, X. Parra, UC Irvine Repository.
"Human Activity Recognition Using
Smartphones," [Online].
https://archive.ics.uci.edu/dataset/240/human+
activity+recognition+using+smartphones
(Accessed Date: January 12, 2024).
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed to the presented
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
This publication has emanated from joint research
conducted with the financial support of the National
Key Research and Development Program of China
under Grant No. 2017YFE0135700 and the
Bulgarian National Science Fund (BNSF) under the
Grant No. KP-06-IP-CHINA/1 (КП-06-ИП-
КИТАЙ/1).
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.4
Xinyi Zeng, Menghua Huang,
Haiyang Zhang, Zhanlin Ji, Ivan Ganchev
E-ISSN: 2224-3402
46
Volume 21, 2024
