Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358656562

Content-based Fashion Recommender System Using Unsupervised Learning

Conference Paper · December 2021


DOI: 10.1109/TENCON54134.2021.9707459

CITATIONS READS

3 458

8 authors, including:

Marielet Guillermo Jason Española


De La Salle University De La Salle University
18 PUBLICATIONS 126 CITATIONS 21 PUBLICATIONS 94 CITATIONS

SEE PROFILE SEE PROFILE

Robert Kerwin Billones Ryan Rhay P. Vicerra


De La Salle University De La Salle University
92 PUBLICATIONS 533 CITATIONS 181 PUBLICATIONS 1,147 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Marielet Guillermo on 03 June 2022.

The user has requested enhancement of the downloaded file.


Content-based Fashion Recommender System
Using Unsupervised Learning
Marielet Guillermo Jason Española Robert Kerwin Billones Ryan Rhay Vicerra
Department of Electronics and Department of Manufacturing Department of Manufacturing Department of Manufacturing
Communications Engineering, Engineering and Management Engineering and Management Engineering and Management
De La Salle University De La Salle University De La Salle University De La Salle University
Manila, Philippines Manila, Philippines Manila, Philippines Manila, Philippines
marielet_guillermo@dlsu.edu.ph jason.espanola@dlsu.edu.ph robert.billones@dlsu.edu.ph ryan.vicerra@dlsu.edu.ph

Argel Bandala Edwin Sybingco Elmer P. Dadios Alexis Fillone


Department of Electronics and Department of Electronics and Department of Manufacturing Department of Transportation
Communications Engineering, Communications Engineering, Engineering and Management and Civil Engineering
De La Salle University De La Salle University De La Salle University De La Salle University
Manila, Philippines Manila, Philippines Manila, Philippines Manila, Philippines
argel.bandala@dlsu.edu.ph edwin.sybingco@dlsu.edu.ph elmer.dadios@dlsu.edu.ph alexis.fillone@dlsu.edu.ph

Abstract— Data mining today is much slower than before As compared to text data, looking for patterns from an
because of the advancement of computing and information image data require more thorough processing and
systems. Relevant recommendation based on customers` mathematical manipulation of an information signal
preferences and needs in e-commerce gets more complicated. In including compression and reconstruction. From performing
the recent pandemic, people are reluctant to go out and has latent-space representation of an image into digits, images are
engaged more on internet to get their daily food and services. reconstructed so that the loss function is at minimum [4]. In
This phenomenon exacerbated the existing recommendation this way, only the most relevant parameters are extracted and
system, as the data has grown up drastically. In this study, the
used to train an unsupervised neural network model. The end
author recommends a relevant image quality based on the
goal is that the digitally processed image has to be almost
quality queries of the clothes and footwear dataset by observing
their highest similarity score. Fashion MNIST images used were
identical to the true image. Consider the example shown in
existing dataset for clothes and footwear. The testing on image Fig. 1 below:
reconstruction using training and validation approaches has
shown an accurate result by showing only 0.01 loss in the
dataset. Using 11 classes of the image queries, the system image
has been identically reconstructed according to the queries
supplied. With this result, businesses will have an
implementation alternative to a faster and more efficient data
mining method. Hence, this alternative will boost the speed of
many recommendation systems in the e-commerce platforms
and will create a better customer experience.
Fig. 1. Convolutional Auto Encoder Architecture
Keywords— image processing, unsupervised neural network,
feature vector, data mining, latent-space representation, digital It can be seen from the image that an auto encoder is
signal processing typically composed of an encoder and a decoder. The encoder
part is responsible for compressing an image dataset and
I. INTRODUCTION
representing it to the latent space while the decoder does the
Big data from aggregated systems opened up new reverse to rebuild the image back to as near as possible to its
opportunities in terms of data mining – opportunity to original appearance. The unsupervised learning method
enhance the existing methods, to speed up the process, and to allows the network model to accommodate more variants of
even jump into the world of computer vision. It is hard to items while cancelling out the problem in data sparsity in the
know the preference of a customer by simply entering queries case of a collaborative recommender system [4]. This is
into a text-based e-commerce search engine [1]. Most often, another type aside from a content-based recommender system
customers may not be aware of the specific text to search for wherein data mining relies on the history of searches or
the products and services they wanted [2]. This complicates interactions [5]. Generally, a cold start problem occurs for
data mining recommendation to work efficiently. Simple startup systems or for newly added items since there is less to
classification methods cannot be applied in this scenario no basis at all to look for recommendations. For a content-
where there are minimal patterns to discern [3]. With this, based system, a single image data query may be enough with
using an unsupervised learning method would be a better fit a well-trained unsupervised convolutional network model.
since it allows discovery of undetermined patterns. More so,
when applied to an image data rather than just a plain text The proponent would implement a content-based fashion
data. As an image can have multiple meanings, it is helpful recommender system with Keras and TensorFlow using a
for a customer to express their search using the unsupervised classless convolutional network model. The results of this
learning method. Hence, the image datasets used in this study. study is aimed to be a starting point towards other
recommender systems in the computer vision field. For
example, a video dataset is being queried. As there is no and sharing of its repetitive parts. Hence, the framework was
model fits all video conditions, specific models may need to called combined candidate network. For easier data
be run for a set of videos. With this study, features can be management and embedding of rich logic functions, data
extracted from a queried video and checked for a similarity were modeled as a Resource Description Framework (RDF)
score with the videos intended for a particular model. graph [9]. In this schema, self-evident truth or rules were
defined for the relationships. TR Discover interface was
II. REVIEW OF RELATED LITERATURE primarily used in this another research study giving users the
Rapid development and integration of software and freedom to enter their query the most natural way [8]. It will
hardware technologies paved a way to people being adaptive be discerned later through the system to make it an executable
and dependent to innovations which offer convenience as query from which an answer can be retrieved accordingly.
oppose to traditional or should we say manual acquisition of Other studies addressed the problem in allocation
products and services [1]. Consequently, vast data has optimization for query processing such as lightweight
entered every system that the problem now is about indexing and hierarchical feature-tree approach in super
optimizing a user experience more than just providing them graph searches [10][11][12].
with an online platform. Some of major bottlenecks in these Abovementioned solutions were more of the data mining
platforms currently are the aggregation and security of per se while there are also these efforts in utilizing deep
information leading to a lower consumer satisfaction [2]. learning for recommender systems. These recommender
Leveraging through a more efficient data mining systems are end user facing. Hence, may dictate user
methodology and making online platforms such as e- retention in an aggregated system depending on how fast and
commerce websites intelligent and convenient contribute to how helpful its search engine produces an output. In the
building an architecture of next generation e-commerce proposed community detection method below, graph
platform [3]. embedding is used for which information on directly
Making queries on a large scale information is rather connected nodes are stored. The graph is then reconstructed,
compute intensive and expensive. Attempts to understand clustered until a given social network is translated into
data flow and to augment data mining processes are in communities [13].
progress [5]. In a study of visualizing data flow graphs, the
aim was to help other researchers to decipher complex
machine learning architectures through visualization so that
the knowledge can be in turn used for making an algorithm
for a much efficient data mining workaround [4].
To solve the problem in coping up with the exponential
increase in data entry, several studies were conducted
especially in database queries. One focused on the quality of
data through categorizing of criteria of interest [6]. It has
objectively defined the term quality to have high rate in
completeness, accuracy, consistency, and freshness. In
Fig. 3. Community Detection Method
reality, query language is way technical to be grasped by an
ordinary application user. Hence, putting up this quality- Among the related solutions presented, the most relevant
aware fuzzy queries so that entering preferences are much to this study is the recommendation model based on user
user-friendly and intuitive. Another study suggested sessions in an e-commerce site [14]. While the objective is
resolutions to tradeoffs in response time and scalability [7]. the same which is to help users find what they are really after
The proponents used divide and conquer approach through from the numerous items available online, this study is more
graph partitions and identified quantitative metrics for of collaborative type. That means historical data is required
heuristics formulation to regulate the stocking division for a more reliable recommendation output. On the other
sequence for query procedural efficiency. Sample algorithm hand, this study is content-based meaning only a single image
used was for a combined candidate network generation. is required and the system will be able to provide a
Framework is illustrated below: recommendation based on its similarity with the rest of
available items. One of the closest study to this is the dataset
recommendation system [15]. Although it is more of text-
based and thus have quite different architecture.

Fig. 2. Combined Candidate Network Evaluation Framework

In a normal setting, each candidate network (CN) is being


handled by relational keyword search which leads to
redundancy [8]. This was solved through compression of CNs Fig. 4. Query-based dataset recommendation. (a) Paper-dataset network;
(b) Dataset recommendation to query
The figure above shows that a pool of research papers vector [20]. The input images are looped over the filters using
with its corresponding datasets were used as an input or actual Keras in the manner below:
dataset network to train a model. This network includes paper
content and citations. As a result of deep learning, the
research paper on query of the best fit dataset is tested and
mapped out to the most relevant and hence recommended
dataset.
III. METHODOLOGY
Querying for a dress image is no different from typing an
Fig. 6. Functional Model
item in a typical search engine. Only, instead of text, an image
is being used. This study is expected to be able to find all In two separate branches above, the images undergo a 3 x
relevant images to a query in a pool of images and return it as 3 convolution with single 2 x 2 stride, decreasing its size. The
an output. To achieve this, the unsupervised learning [16][17]
next branch performs max-pooling with the same size and
of a queried image is used following the framework below:
stride. At the end, the output from both branches are merged
across the channel dimension. After passing through the loop,
a batch normalization is applied and ends up with a flattened
network. The dense layer now will serve as the latent-space
representation of a given image and will be stored for
comparison on the later part of the block diagram. For the
reconstruction of the image, filters are looped over but on a
reverse order and with application of a transposed
convolution. This layer is used to recover the original spatial
dimensions [21]. Applying a sigmoid activation function, the
values are transformed to between 0.0 and 1.0 only, a way to
quantify contents of a given image [22].
In training the network model, Adam optimization
Fig. 5. Block Diagram algorithm was used to manage issues in unwanted signals. It
is relatively a good alternative for the traditional stochastic
First, image datasets for each of the 10 fashion categories gradient descent process in terms of updating network
need to be made available and searchable in the system.
weights iteration in accordance with the training data
Fashion MNIST images used are 28 pixels both in height and
[23][24]. For this study, the hyper parameter constants used
in width with a single pixel value associated and in grayscale
form. The training set is composed of over 60, 000 samples are 20 for epoch, 0.001 for the initial learning rate, and 32 for
and 10, 000 for the test set. These have to be loaded together the batch size [25][26]. Once training of the network model
with the network model. Through unsupervised learning [31][32] is done and an image is queried, a comparator is
method, the network model will be trained with the latent- triggered through a helper function which has the ability to
space vector representations for each of the images [18]. The do predictions visually. Inside this function, ground truth
encoder part of the network is used to compute for the feature images are being compared to the reconstructed output of the
vector of an image which quantifies its contents [19]. Given network model and looped over the available stacked images.
that E represents the encoder, i for the image, and latent-space
The similarity between images are being calculated
representation is l then:
through the Euclidean distance between its feature vectors.
l = E(i) (1) The working formula is below:
After successful quantification of image content, d2 (x,y,) = (x-y)T (x-y) = xTx+yTy-2xTy
reconstruction from the space representation to original input
follows. At this stage, the mean squared error or the similar = |x|+|y|-2[cos(x,y) x |x| x |y|] (4)
loss function is minimized. With the decoder function = ∑(xi-yi) 2
denoted as D and the output image, o:
The x and y variable in the formula shown refer to the 1 st
o = D(l) (2) and 2nd vector while d stands for the resulting distance.
With these, the full training procedure can be Among several methods of computing distance vectors
characterized in the mathematical equation: [27][28], Euclidean was chosen as it can translate minimum
distance between data points which do not belong from the
o = D(E(i)) (3) same cluster while it can maximize if from different clusters
From a total of 784 pixels, image pixel value is [29]. The figure below illustrates sample data points and its
compressed to a vector value of less than 20. Although the corresponding cluster center.
output image is typically a replica of the input but with a
compressed value, any outlier from a typical data distribution
can be immediately cancelled out.
Five parameters namely height, width, filters, depth, and
latentDim, are considered in the construction of the latent
well as the validation loss when the epoch reaches closer to
its last value. An epoch of 20 was used for practicality
although there’s a lot of datasets available [33]. The said
value is more than enough but can also be increased
depending on data variation and goal of the network model to
be trained. The graph below serves as a visualization of the
training and validation loss data behavior.

Fig. 7. Two Clusters in Two-Dimensional Space

The closer the data points to a cluster center means it is


more likely that this data point belong to that group/ cluster.
On the other hand, the farther the data point to a cluster center
means the probability that it belongs to that group is lower.
Applying it to this study, the smaller the Euclidean distance
result means two images are most likely similar [30]. Images
are then different if the distance is too high. To test the
reliability of the system, some of the images from Fashion Fig. 9. Training Loss and Accuracy
MNIST dataset were used as test sets and are randomly
picked. The recommended images based on similarity are Below is the result of the reconstruction of images. It can
then shown in a montage form. This essentially now is the be seen that the replica is almost the same as the input which
content-based fashion recommender system. Based in the implies that the network model did well in latent-space
input or queried image, the system will search from the stored representation of image datasets. The classes shown are only
images and show it as an output. For this study, maximum 10, although, there are more than available from datasets that
number of image results that can be shown is 225. The results were reconstructed and stored in the system.
in terms of training time, search speed, accuracy, and other
important parameters will be discussed in detail in the next
section.

IV. DATA AND RESULTS


Training of the network model was run in an Ubuntu
machine with a computing-related specifications:
• Processor: Intel® Core™ i7-8565U CPU @ 1.80
GHz
• CPU(s): 8
• Architecture: x86_64
Fig. 10. Reconstructed Images
With the hyperparameters indicated in the previous
section, the training lasted for about 23 minutes starting with This result is an important part of making
a training and validation loss of 0.0180 and 0.0131 recommendations. Since the output is precise, the outcome of
respectively. For each epoch, the average time is 90 seconds the fashion recommendation system as shown below is
which can be translated to approximately 48 milliseconds per accurate enough. The next figures are from 11 classes
sample. available in the Fashion-MNIST dataset and its
corresponding recommended similar fashion images.

Fig. 8. Network Model Training

The figure above shows the training performance per


epoch. It can be seen that the training loss is diminishing as Fig. 11. Class 1 and 2
Fig. 12. Class 3 and 4

Fig. 16. Class 11

The classes above vary from shoes, top clothes, and pants.
All are fashion related images. With the 11 random queries,
similar images to it which are stored in the system were
shown in the Results section.
V. CONCLUSION AND RECOMMENDATIONS
Recommender systems are becoming a necessity
nowadays due to the fact that most software applications are
being aggregated, providing a one stop shop for customers
especially in the field of fashion. With the data and results of
the study discussed from the previous section, it can be
concluded that the demonstration of an unsupervised learning
of images was realized. The feature extraction and
Fig. 13. Class 5 and 6 reconstruction are also excellent. Hence, recommendation
output of similar fashion images per query was accurate and
fast. With this as the starting point of other recommender
systems and data mining processes, it is indeed feasible to
provide suggestions despite the cold start or lack of history of
transactions especially for first time users if an application.
Aside from feasibility, aggregated systems will be more
efficient and user friendly in a sense that a user does not need
to go through the entire categories of an online shop or even
to search from different applications, just to know and search
for the item that is being looked for. Not only users can save
time in online shopping, but also, they can have the assurance
of getting all the suggestions relevant to his/her item query.
As a recommendation, other researchers may want to
implement this study with a live or actual e-commerce
website. They can also try to explore its implementation in
Fig. 14. Class 7 and 8 other online systems like a delivery service, streaming
services, and even in other machine learning-based systems.
On another note, they can also consider other factors aside
from similarity such as rating, price, brand, and even the
seller of an item. In terms of the network model, it is
recommended to explore other algorithms or train with other
datasets to further strengthen its robustness, boost its
accuracy, and improve its speed.
ACKNOWLEDGMENT
De La Salle University (DLSU) Intelligent Systems
Laboratory (ISL) is highly appreciated for the resources and
opportunity to broaden the knowledge and experience of the
researcher. The Department of Electronics and
Communications Engineering is equally thanked for the
motivation in venturing to a deeper research field. With the
Fig. 15. Class 9 and 10 guidance of its credible and highly experienced professors,
the researcher was able to widen her innovative ideas and [18] W. Didimo, L. Grilli, G. Liotta, F. Montecchiani, and D. Pagliuca,
enhance her critical thinking. Lastly, Department of Science “Visual querying and analysis of temporal fiscal networks,” Inf. Sci.
(Ny)., vol. 505, pp. 406–421, 2019, doi: 10.1016/j.ins.2019.07.097.
and Technology - Philippine Council for Industry, Energy
[19] Z. Cui, K. Henrickson, R. Ke, and Y. Wang, “Traffic Graph
and Emerging Technology Research and Development Convolutional Recurrent Neural Network: A Deep Learning Framework
(DOST-PCIEERD) is hereby acknowledged for the for Network-Scale Traffic Learning and Forecasting,” IEEE Trans.
scholarship and financial grant, enabling the researcher to get Intell. Transp. Syst., pp. 1–12, 2019, doi: 10.1109/tits.2019.2950416.
an advanced research experience. [20] L. Chen, Y. Xie, Z. Zheng, H. Zheng, and J. Xie, “Friend
Recommendation Based on Multi-Social Graph Convolutional
REFERENCES Network,” IEEE Access, vol. 8, pp. 43618–43629, 2020, doi:
10.1109/ACCESS.2020.2977407.
[1] X. Yang, “Satisfaction evaluation and optimization of tourism e-
commerce users based on artificial intelligence technology,” Proc. - [21] K. Guo et al., “Optimized Graph Convolution Recurrent Neural
2019 Int. Conf. Robot. Intell. Syst. ICRIS 2019, pp. 373–375, 2019, doi: Network for Traffic Prediction,” IEEE Trans. Intell. Transp. Syst., pp.
10.1109/ICRIS.2019.00100. 1–12, 2020, doi: 10.1109/tits.2019.2963722.
[2] Q. Wu, J. Ma, and Z. Wu, “Consumer-driven e-commerce: A study on [22] Z. Zhang, J. Leng, L. Ma, Y. Miao, C. Li, and M. Guo, “Architectural
C2B applications,” Proc. - 2020 Int. Conf. E-Commerce Internet Implications of Graph Neural Networks,” IEEE Comput. Archit. Lett.,
Technol. ECIT 2020, pp. 50–53, 2020, doi: vol. 19, no. 1, pp. 59–62, 2020, doi: 10.1109/LCA.2020.2988991.
10.1109/ECIT50008.2020.00019. [23] Z. Chu, J. Yu, and A. Hamdulla, “A novel deep learning method for
[3] Y. Huang, Y. Chai, Y. Liu, and J. Shen, “Architecture of next-generation query task execution time prediction in graph database,” Futur. Gener.
e-commerce platform,” Tsinghua Sci. Technol., vol. 24, no. 1, pp. 18– Comput. Syst., vol. 112, pp. 534–548, 2020, doi:
29, 2019, doi: 10.26599/TST.2018.9010067. 10.1016/j.future.2020.06.006.
[4] K. Wongsuphasawat et al., “Visualizing Dataflow Graphs of Deep [24] T. Bogaerts, A. D. Masegosa, J. S. Angarita-Zapata, E. Onieva, and P.
Learning Models in TensorFlow,” IEEE Trans. Vis. Comput. Graph., Hellinckx, “A graph CNN-LSTM neural network for short and long-
vol. 24, no. 1, pp. 1–12, 2018, doi: 10.1109/TVCG.2017.2744878. term traffic forecasting based on trajectory data,” Transp. Res. Part C
[5] Q. Song, Y. Wu, P. Lin, L. X. Dong, and H. Sun, “Mining Summaries Emerg. Technol., vol. 112, no. December 2019, pp. 62–77, 2020, doi:
for Knowledge Graph Search,” IEEE Trans. Knowl. Data Eng., vol. 30, 10.1016/j.trc.2020.01.010.
no. 10, pp. 1887–1900, 2018, doi: 10.1109/TKDE.2018.2807442. [25] M. Guillermo et al., “Detection and Classification of Public Security
[6] O. Pivert, E. Scholly, G. Smits, and V. Thion, “Fuzzy quality-Aware Threats in the Philippines Using Neural Networks,” 2020, doi:
10.1109/LifeTech48969.2020.1570619075.
queries to graph databases,” Inf. Sci. (Ny)., vol. 521, pp. 160–173, 2020,
doi: 10.1016/j.ins.2020.02.035. [26] M. Guillermo et al., “Implementation of Automated Annotation through
[7] S. Das, A. Santra, J. Bodra, and S. Chakravarthy, “Query processing on Mask RCNN Object Detection model in CVAT using AWS EC2
Instance,” 2020, doi: 10.1109/tencon50793.2020.9293906.
large graphs: Approaches to scalability and response time trade offs,”
Data Knowl. Eng., vol. 126, no. September 2019, p. 101736, 2020, doi: [27] M. C. Pabón, M. Millán, C. Roncancio, and C. A. Collazos, “GraphTQL:
10.1016/j.datak.2019.101736. A visual query system for graph databases,” J. Comput. Lang., vol. 51,
no. April, pp. 97–111, 2019, doi: 10.1016/j.cola.2018.12.006.
[8] G. Ding, H. Sun, J. Li, C. Li, R. Wei, and Y. Fei, “An Efficient
Relational Database Keyword Search Scheme Based on Combined [28] S. Xu, E. S. L. Ho, N. Aslam, and H. P. H. Shum, “Unsupervised
Candidate Network Evaluation,” IEEE Access, vol. 8, pp. 30863–30872, abnormal behaviour detection with overhead crowd video,” Int. Conf.
2020, doi: 10.1109/ACCESS.2020.2973217. Software, Knowl. Information, Ind. Manag. Appl. Ski., vol. 2017-
Decem, pp. 0–5, 2018, doi: 10.1109/SKIMA.2017.8294092.
[9] D. Song et al., “Building and Querying an Enterprise Knowledge
Graph,” IEEE Trans. Serv. Comput., vol. 12, no. 3, pp. 356–369, 2019, [29] D. Mu, L. Guo, X. Cai, and F. Hao, “Query-Focused Personalized
doi: 10.1109/TSC.2017.2711600. Citation Recommendation with Mutually Reinforced Ranking,” IEEE
[10] K. Lee et al., “Lightweight Indexing and Querying Services for Big Access, vol. 6, pp. 3107–3119, 2017, doi:
10.1109/ACCESS.2017.2787179.
Spatial Data,” IEEE Trans. Serv. Comput., vol. 12, no. 3, pp. 343–355,
2019, doi: 10.1109/TSC.2016.2637332. [30] Z. Wang, Y. Yuan, L. Chang, X. Sun, and X. Luo, “A Graph-Based
[11] B. Lyu, L. Qin, X. Lin, L. Chang, and J. X. Yu, “Supergraph Search in Visual Query Method for Massive Human Trajectory Data,” IEEE
Graph Databases via Hierarchical Feature-Tree,” IEEE Trans. Knowl. Access, vol. 7, pp. 160879–160888, 2019, doi:
10.1109/ACCESS.2019.2948304.
Data Eng., vol. 31, no. 2, pp. 385–400, 2019, doi:
10.1109/TKDE.2018.2833124. [31]R. K. C. Billones et al., "Vision-Based Passenger Activity Analysis
[12] A. B. Mathew, “Data allocation optimization for query processing in System in Public Transport and Bus Stop Areas," 2018 IEEE 10th
graph databases using Lucene,” Comput. Electr. Eng., vol. 70, pp. 1019– International Conference on Humanoid, Nanotechnology, Information
1033, 2018, doi: 10.1016/j.compeleceng.2018.01.022. Technology,Communication and Control, Environment and
Management (HNICEM), 2018, pp. 1-6, doi:
[13] L. Wu, Q. Zhang, C. H. Chen, K. Guo, and D. Wang, “Deep learning 10.1109/HNICEM.2018.8666357.
techniques for community detection in social networks,” IEEE Access,
vol. 8, pp. 96016–96026, 2020, doi: 10.1109/ACCESS.2020.2996001. [32]I. J. C. Valencia, E. P. Dadios, A. M. Fillone, J. C. V. Puno, R. G.
Baldovino and R. K. C. Billones, "Vision-based Crowd Counting and
[14] H. Wang, G. Xiao, N. Han, and H. Chen, “Session-based graph Social Distancing Monitoring using Tiny-YOLOv4 and DeepSORT,"
convolutional ARMA filter recommendation model,” IEEE Access, vol. 2021 IEEE International Smart Cities Conference (ISC2), 2021, pp. 1-
8, pp. 62053–62064, 2020, doi: 10.1109/ACCESS.2020.2984039. 7, doi: 10.1109/ISC253183.2021.9562868.
[15] B. Altaf, U. Akujuobi, L. Yu, and X. Zhang, “Dataset recommendation [33]R. K. C. Billones et al., "Vehicle-Pedestrian Classification with Road
via variational graph autoencoder,” Proc. - IEEE Int. Conf. Data Mining, Context Recognition Using Convolutional Neural Networks," 2018
ICDM, vol. 2019-Novem, no. Icdm, pp. 11–20, 2019, doi: IEEE 10th International Conference on Humanoid, Nanotechnology,
10.1109/ICDM.2019.00011. Information Technology,Communication and Control, Environment and
[16] H. Shi et al., “Hypergraph-Induced Convolutional Networks for Visual Management (HNICEM), 2018, pp. 1-6, doi:
Classification,” IEEE Trans. Neural Networks Learn. Syst., vol. 30, no. 10.1109/HNICEM.2018.8666257
10, pp. 2963–2972, 2019, doi: 10.1109/TNNLS.2018.2869747.
[17] C. Li, X. Qin, X. Xu, D. Yang, and G. Wei, “Scalable Graph
Convolutional Networks With Fast Localized Spectral Filter for
Directed Graphs,” IEEE Access, vol. 8, pp. 105634–105644, 2020, doi:
10.1109/access.2020.2999520.

View publication stats

You might also like