This study aims to explore new methods of website creation to meet the surging demand
for website production, spurred by the increasing demand for online businesses and the
importance of non-face-to-face services. It has designed and implemented a sys...
This study aims to explore new methods of website creation to meet the surging demand
for website production, spurred by the increasing demand for online businesses and the
importance of non-face-to-face services. It has designed and implemented a system capable
of generating actual websites based on artificial intelligence and data, technologies of the
Fourth Industrial Revolution. Usability evaluations were conducted with IT experts and
general users to validate the proposed system.
The study was carried out in the following steps: First, it researched the overall process
and environment of website production, exploring the evolution of technology and
production environments, and identifying limitations. It examined the industry from website
production agencies to Wix's ADI (Artificial Design Intelligence) service and sought ways
to apply deep learning to automatic website generation systems. Second, it reviewed prior
research on artificial intelligence technologies, identified necessary deep learning models,
and selected models suitable for the analysis and inference of websites. Third, it established
a classification system based on industry types and functions of websites, classified
paragraphs and components of web pages, and designed a complete automatic website
generation system applicable to deep learning. Finally, it trained four types of deep learning
models, implemented inference capabilities, and integrated them with the automatic website
generation system to implement the full functionality of the system. Usability evaluations
were conducted with experts and the general public to verify the system's usability, aiming
to improve its effectiveness and directions for development.
The general procedure for website production consists of data collection, planning,
design, and development. This process has not changed significantly since the early 2000s
when the web spread rapidly in Korea, although there have been significant changes in the
production and service environments. The most significant change is that the production
environment is now mostly web-based. Whereas past web utilization was limited to email
or file-sharing services, recent practices involve using collaborative planning tools like
Google Sheets and presentations and collaborative design tools like Figma on websites.
Since the 2010s, services known as automatic website generation systems have been rapidly
disseminated, enabling even development to be completed on a web basis. These systems
allow for the easy implementation of various functions necessary for development with a
selection of features and some settings. Furthermore, the range of services these systems can
provide is gradually expanding, now encompassing planning and design.
This shift in the production environment paradigm towards automatic website
generation systems has led to the rapid emergence of the 'No-Code' production method,
which does not require development. In the field of design, the concept of 'Computational
Design,' a new concept based on computing and cutting-edge technology, is expanding. This
paradigm shift is changing to quickly produce cohesive and empathetic designs based on
templates. Since this paradigm shift is all based on web systems, it enables the collection
and automation of production data and procedures.
The core technology for automatically generating websites lies in the learning and
inference of artificial intelligence models. Therefore, to use the most suitable artificial
intelligence technology, a study was conducted on artificial intelligence technologies.
According to Aurelien Geron's book 'Hands-On Machine Learn,' artificial intelligence
technologies mainly consist of supervised learning, unsupervised learning, and
reinforcement learning, each with different usage areas, technical characteristics, and
application fields. This study falls under supervised learning, as it is based on existing data
for learning and finding answers. Deep learning in supervised learning includes three major
types: classification, object detection, and image segmentation. The study deemed a
'classification algorithm' for analyzing images and an 'algorithm for detecting elements' as
necessary.
To construct an environment where artificial intelligence models can learn and infer, it
is necessary to collect and process vast amounts of website data. While it was difficult to
collect numerous industry-specific and function-specific website production data in the past,
it is now possible to collect diverse and extensive data due to the proliferation of automatic
website generation systems, with the entire production procedure being web-based. This
study utilized an automatic website generation system to collect and process data,
conducting learning and inference of artificial intelligence models. The study also designed
and implemented an automatic website generation system based on artificial intelligence and
its integration with the web.
Next, the environment for the artificial intelligence-based automatic website generation
service must be a system capable of stably storing and using large amounts of data and
providing stable service against the traffic of numerous users. Therefore, the infrastructure
environment was determined to be cloud-based and designed as a SaaS (Software as a
Service) according to Sether Ayob's advantages of cloud computing (Ayob 2016).
The most prioritized part of the website production procedure is to understand the
minimum classification and functional requirements of the industry to which the customer's
desired website belongs. For this, the simplest and most effective method was deemed to be
receiving the address or image of a benchmarking website. When a website address is
entered, the system was designed to download the website as an image using the scraping
function proposed by Moaiad Ahmad Khder (Khder 2021). Then, the deep learning model
analyzes the image, and the inference results are matched with the data of the automatic
website generation system, following the overall structure and flow of the automatic code
generation (ACG) process proposed by John Vlissides (Budinsky, Finnie et al. 1996). The
design of artificial intelligence involved implementing at least three types of models for
analyzing and inferring websites. The first model, YOLOV8, was deemed most suitable for
'detecting paragraphs' that make up a page, using a single-layer object detection algorithm.
The second model, EfficientNetV2, was deemed suitable for 'classifying the types and
numbers of paragraphs' using a convolutional neural network's image classification
algorithm. The third model, similar to the first, was deemed suitable for 'detecting elements
such as images or text within paragraphs.' Subsequently, the study collected and labeled
about 1,000 e-commerce web pages for paragraphs, types, and elements, and based on this
data, trained the models. The trained models were then implemented in a serving
environment, and the entire interface capable of analysis and inference was implemented.
Finally, the study implemented 'DWAGS (Deep Learning-Based Website Automatic
Generation System),' a system that can automatically generate actual websites by receiving
the address or image of a benchmarking website.
To verify whether websites generated by DWAGS are suitable for immediate use in the
industry, a survey was conducted among actual IT experts and the general public. The first
evaluation yielded an average score of 3.8, confirming its usability, but various opinions for
usability improvement were converged, leading to the improvement of DWAGS. A second
evaluation was conducted with more survey participants to track the results. The second
usability evaluation resulted in an average score of 4.2, a 10% increase in evaluation scores,
and particularly high praise from experts, confirming its applicability in the industry.
Previous studies and cases of services provided by companies show that while research
and development in artificial intelligence are very active, image-based results are often
awkward or do not meet user intentions, making them insufficient for immediate application
in the industry. However, this study is of high practical value as it concretely proposes a new
suggestion that can be used in the industry by completely redesigning the website automatic
generation system based on data, leveraging the innovative technologies of artificial
intelligence, a leading technology of the Fourth Industrial Revolution. Although research in
the field of artificial intelligence-based image generation and design is actively underway,
the results are still fragmented and have clear limitations. Therefore, this study is significant
in the industry as a new data-based automatic generation system applying deep learning and
academically as a representative example of how innovative services can be integrated into
actual industries. Moreover, by concretizing the classification system for various paragraphs
and elements, which was previously lacking from an industry-specific and functional
perspective, the study contributes to the knowledge of various web industries, such as feature
development. Furthermore, it is expected to create a new paradigm in website production
and contribute to the development of the industry by providing fast and convenient services
to various entrepreneurs aspiring to online businesses, based on the maximization of
productivity with artificial intelligence.