Earticle

폐렴은 폐에 고름 혹은 물이 차는 호흡기 관련 감염 질환이다. 세계적으로 폐렴은 소아 및 노인 계층에서 감염 질환 으로 인한 사망 원인 중 하나이다. 일반적으로 의료진들은 폐렴 진단을 위해 흉부 X-ray 영상을 사용한다. 그러나 폐렴은 유사한 증상을 보이는 감기와 독감으로 오진될 가능성이 높으며, 이는 심각한 합병증의 원인이 될 수 있다. 따라서, 의료진의 실수를 최소화하기 위해 대안적인 보조 진단 방법이 필요하다. 폐렴의 조기 진단을 위한 보조 도 구로 딥러닝 기술에 기반한 인공지능 시스템이 연구되고 있으나, 딥러닝이 진단 보조 도구로 확립하기 위해서는 더 많은 연구가 필요하다. 본 연구에서는 광저우 여성 아동 병원 데이터셋과 코로나19·폐렴·정상 흉부 X-ray 데이터 셋을 사용하여 Swin Transformer와 Residual Neural Network의 성능을 분석한다. 실험 결과, Swin Transformer는 흉부 X-ray 영상 데이터셋에서 98.9%의 정확도를 보이며,코로나19·폐렴·정상 흉부 X-ray 데이 터셋에서는 92.35%의 정확도를 보인다. Residual Neural Network는 두 데이터셋에서 각각 97.9%와 88.8% 의 정확도를 보인다. 이러한 결과는 폐렴 보조 진단 도구로써 Swin Transformer가 Residual Neural Network 보다 적합함을 의미한다. 그러므로, Swin Transformer는 조기 진단과 조기 치료, 환자의 건강 증진에 기여할 수 있을 것이라 기대한다.

PPneumonia is a respiratory infectious disease that causes fluids to fill the lungs. It is considered one of the leading causes of infection-related deaths in children and seniors worldwide. Clinicians usually use chest X-ray images to diagnose pneumonia. However, pneumonia is prone to be misdiagnosed because it overlaps with cold and flu, causing severe and critical medical complications. Consequently, alternative supportive diagnostic methods are needed to minimize human errors and assist clinicians. Several attempts have used artificial intelligence systems, mainly in deep learning methods, to assist clinicians in early pneumonia diagnoses. However, further studies are required to consolidate the use of deep learning as an assistant tool to diagnose pneumonia accurately. In this study, we examine the Swin Transformer and the Residual Neural Network’s performance in classifying pneumonia and healthy chest X-ray images using the Guangzhou Women and Children’s Medical Center dataset and the COVID19, Pneumonia and Normal Chest X-ray Posteroanterior dataset. The experiment results demonstrate that the Swin Transformer achieves an accuracy of 98.9% in the Chest X-ray images dataset and 92.35% in the COVID19, Pneumonia and Normal Chest X-ray Posteroanterior dataset, while the Residual Neural Network achieves an accuracy of 97.9% and 88.8% respectively in classifying pneumonia. These results indicate that the Swin Transformer outperforms the Residual Neural Network as a tool for assisting clinicians in diagnosing pneumonia. Thus, the Swin Transformer may help in early decision-making, leading to treatment initiation and improving patient's health.

Customer reviews for wireless earbuds were collected and preprocessed using Playwright and Requests-HTML libraries, ensuring high-quality and relevant data. This paper introduces sentiments associated with these aspects were identified using Recurrent Neural Networks (RNNs) and Bidirectional Encoder Representations from Transformers (BERT) enhanced with attention mechanisms, which helped focus on the most relevant text segments. The models were integrated using ensemble methods, specifically Voting+BERT and Bagging+BERT, to improve accuracy and robustness. The Bagging+BERT model achieved the best performance, with an accuracy of 89.9 %, outperforming traditional machine learning models like Bayesian and logistic regression by 9.6 % and 8.7 %, respectively.

Sepsis and Neonatal sepsis are major challenges in global healthcare because they cause life-threatening organ dysfunction in intensive care adult and pediatric patients due to downregulated host response to a particular infection. Early clinical identification of sepsis is difficult, and failure to provide prompt treatment can often lead to crucial stages and increase the rates of fatality. Thus an intense study is needed to determine and categorize sepsis in its initial stage. The complexity of varying clinical statistics makes it difficult to attain a precise definition in pediatrics. The advanced Machine Learning (ML) and Deep Learning (DL) technologies in the implementation of protocols show promising real-time models for predicting sepsis at the primary stage and thereby reducing the mortality rate. This review article contemplates the complete list of procedures through which sepsis and neonatal sepsis are speculated by ML and DL and concentrates specifically on data available in the adult emergency care unit as well as the neonatal intensive care unit. The survey process was carried out by searching terms related to ML and DL merged with topics concerning sepsis and neonatal sepsis. The literature analysis was carried out from Scopus, Web of Science, and PubMed databases for the period from 2015 to 2022. The assessment of the risk of bias was carried out for the eleven selected papers using the Prediction Model Risk of Bias Assessment Tool (PROBAST). The eleven papers were selected from different medical care units based on the performance measure AUROC, which ranges from 0.68 to 0.95. Five papers involving ML/DL models reduce the bias and lessen risk occurrence. Five papers generate an increase in bias but can be applied to new data. One paper works with above twenty-five features has high-risk probability but predicts patients within 5?6 h in the future. This survey portrays the role of prediction models that supports the researchers and clinicians for better decision-making and antibiotic administration at an earlier stage.

Although deep learning (DL) has already become a state-of-the-art technology for various data processing tasks, data security and computational overload problems often arise due to their high data and computational power dependency. To solve this problem, quantum deep learning (QDL) and distributed deep learning (DDL) has emerged to complement existing DL methods. Furthermore, a quantum distributed deep learning (QDDL) technique that combines and maximizes these advantages is getting attention. This paper compares several model structures for QDDL and discusses their possibilities and limitations to leverage QDDL for some representative application scenarios.

인터넷 서비스의 확산과 함께 정교해지는 피싱(Phishing) 공격은 개인 정보 탈취 및 금융 피해를 유발하는 심각한 보안 위 협으로 대두되고 있다. 기존의 피싱 탐지 체계는 주로 구글 세이프 브라우징(Google Safe Browsing)이나 피쉬탱크 (PhishTank)와 같은 블랙리스트(Blacklist) 방식에 의존해 왔다. 이 방식은 알려진 위협에 대해서는 신속하고 정확한 차단이 가능하나, 제로데이(Zero-day) 공격을 탐지하지 못하는 치명적인 한계를 가진다. 본 연구에서는 이러한 한계를 극복하기 위해 URL의 어휘적 특징을 기반으로 하는 다양한 인공지능 모델의 탐지 성능을 비교 분석하였다. 실험 대상 모델로는 전통적인 휴 리스틱 알고리즘과 머신러닝 모델인 로지스틱 회귀(Logistic Regression), 서포트 벡터 머신(SVM), 랜덤 포레스트(Random Forest), 그리고 딥러닝 모델인 CNN(1D)과 LSTM을 선정하였다. 실험 결과, 휴리스틱 방식은 44.5%의 저조한 정확도를 보인 반면, SVM(RBF 커널) 모델은 97.0%의 정확도와 0.970의 F1-Score를 기록하며 가장 우수한 성능을 나타냈다. 특히 딥러닝 모 델인 CNN(94.5%)과 LSTM(76.1%) 대비 SVM은 0.165초라는 빠른 추론 속도를 보여 실시간 탐지 환경에서 성능과 효율성의 최적 균형을 갖춘 모델임을 입증하였다.

As internet services proliferate, phishing attacks are becoming increasingly sophisticated and are emerging as a serious security threat that causes the theft of personal information and financial damage. Existing phishing detection systems have primarily relied on blacklist methods such as Google Safe Browsing or PhishTank. While this approach enables the rapid and accurate blocking of known threats, it has a critical limitation in its inability to detect zero-day attacks. To overcome these limitations, this study comparatively analyzed the detection performance of various artificial intelligence models based on the lexical features of URLs. The models selected for the experiment included traditional heuristic algorithms, machine learning models such as Logistic Regression, Support Vector Machine(SVM), and Random Forest, as well as deep learning models like CNN(1D) and LSTM. The experimental results showed that while the heuristic method yielded a poor accuracy of 44.5%, the SVM(RBF kernel) model demonstrated the superior performance, recording an accuracy of 97.0% and an F1-Score of 0.970. In particular, compared to the deep learning models CNN(94.5%) and LSTM(76.4%), SVM demonstrated a fast inference speed of 0.165 seconds, proving it to be the model with the optimal balance between performance and efficiency in a real-time detection environment.

2004년 오염총량제 도입 이후, 하천 유량자료의 중요성이 대두되었다. 그러나 총량 모니터링 지점은 8일 주기로만 측정을 수행하기 때문에, 연속적이고 신뢰성 있는 유량자료를 제공하지 못하는 실정이다. 이에 반해 하류의 유량관측소는 수위–유량 곡선식을 활용하여 연속적인 측정을 수행하고 있다. 본 연구 에서는 총량 모니터링 지점의 유량 산정을 개선하기 위해 강우–유출 모형인 GR4J와 딥러닝 모형인 LSTM을 적용하였다. 총량 모니터링 지점의 결과를 유량관측소에서의 유량 산정 결과와 비교하기 위해 2013년부터 2023년까지의 자료를 활용하였으며, 그 결과 두 지점 간의 상관계수는 0.91 이상으로 강한 상관성을 보였다. 예측 성능 측면에서 LSTM 모형은 GR4J 모형보다 더 우수한 결과를 보였다. 특히 유량관측소 자료와 GR4J 모형의 시뮬레이션 결과를 함께 활용했을 때, 훈련 단계에서 상관계수 0.99, 검증 단계에서 0.95라는 안정적인 성능을 보였다. 본 연구의 결과는 단속적 측정으로 인한 유량자료의 불연속성을 해결하는 기술적 방법을 제시하며, 수자원 공급과 총량제 운영의 신뢰성을 높이는 데 기여할 것으로 사료된다.

Streamflow data have become more important after the introduction of the Total Pollution Load Management System in 2004. However, Total Pollution Load Monitoring Stations record measurements every 8 days and cannot provide continuous and reliable flow data. In contrast, Flow Monitoring Stations provide continuous data using rating curves. In this study, the GR4J (Ge'nie Rural a 4 parameters Journalier) Rainfall–Runoff Model and the Long Short-Term Memory (LSTM) Model were used to improve flow estimation from Total Pollution Load Monitoring Stations. From 2013 to 2023, results of Total Pollution Load Monitoring Stations were compared with flow estimates from Flow Monitoring Stations, showing a high correlation with a correlation coefficient greater than 0.91. The LSTM Model performed better than the GR4J model. Flow Monitoring Station data combined with GR4J simulations produced stable results, with correlation coefficients of 0.99 (training) and 0.95 (testing). In conclusion, the results provide a practical method to fill voids caused by intermittent measurements and improve reliability in water supplies and total load management.

Skin cancer, particularly melanoma, poses significant risks due to its high metastatic potential and challenges in early diagnosis. Accurately detecting skin lesions through automated systems is crucial for improving survival rates. This paper does not merely propose a detection method but analyzes the effectiveness of feature extraction for accurate skin lesion classification. Utilizing a dataset from Kaggle, this paper compares the performance of various deep learning models, including Convolutional Neural Networks (CNN), Deep Neural Networks (DNN), and ResNet-18. We evaluate the ability to classify skin lesions by training three models on 10,015 images across seven classes. ResNet-18 achieved the highest accuracy of 81.6%, demonstrating its potential for the development of automated diagnostic systems. In contrast, CNN and DNN attained lower accuracies of 72.9% and 70%, respectively, likely due to limitations in their feature extraction capabilities. These results underscore the superior performance of ResNet-18, particularly in its ability to handle complex patterns and deep feature learning, which are critical for skin lesion classification. In addition, we explored the potential integration of Large Language Model(LLM) to enhance the interpretability of diagnostic outcomes. By utilizing the Llama2 model API provided by Hugging Face, we explained the feasibility of interpreting ResNet-18's predictions to provide users with more transparent and higher-level medical insights. This suggests a promising future direction for improving the explainability and clinical applicability of AI-driven skin lesion diagnosis.

This study emphasizes the necessity of artificial intelligence for rapid and accurate battery state-of-charge (SOC) prediction, a critical parameter in battery condition prediction. We compared and evaluated time series models previously used for SOC prediction, namely LSTM, GRU, and Transformer. In addition to model comparison, we experimented with data preprocessing techniques suitable for battery SOC prediction. The study utilized NASA's aging dataset comprising different cells under various experimental conditions. A Sliding Window technique was employed to multiply data and evaluate model performance. The results showed that the GRU model most effectively predicted battery SOC without data multiplication. However, after applying the Sliding Window technique to generate more learning data, the Transformer model outperformed others with an average RMSE of 0.032 and MAE of 0.006 across all batteries. This research paves the way for advancements in AI technology based on Transformer models for improved analysis of battery conditions, which can benefit manufacturing and recycling processes.

Transfer learning significantly improves the performance of a deep learning model on challenging datasets. However, the pre-trained models have certain constraints in terms of their architecture. For example, the state-of-the-art pre-trained models expect an input image with three-color channels because of the wide availability of color images. However, there are certain domains, e.g., medical applications, where grayscale images are produced and the models are required to perform certain tasks on them. Therefore, in this work we propose an approach to run pre-trained models on grayscale images while benefiting from transfer learning for multiclass classification task. We have used the MobileNetV2 pre-trained model to classify the CIFAR datasets. We have compared our results with a conventional method where the grayscale image is stacked up to form a pseudo-color image. Our analysis have shown that the proposed method reduces the computational time per epoch while improves the accuracy of the model.

As there is no cure for fire blight, which mainly affects pears and apples, effective and rapid detection is very important. Existing fire blight diagnostic studies usually used biotechnology, such as immunodiagnostic kits. With the development of deep learning-based image recognition technology, an image-based fire blight diagnosis method has been proposed. For the diagnosis of diseases that have similar symptoms, including fire blight, this study developed a disease recognition model using the deep convolutional neural network (CNN). Fine-tuning was performed on VGG16, VGG19, ResNet50, DenseNet121, Inception-ResNet v2, NASNet and EfficientNet models, which were pre-trained through ImageNet dataset. The experiment used 14,304 images of six diseases collected from pear and apple as the dataset. As a result of the experiment, all seven fine-tuned models achieved an accuracy of more than 90%, among which the ResNet50 model achieved the highest accuracy at 98.83%. It is anticipated that the proposed model can be valuably used at actual farmhouses to diagnose and prevent fire blight through appropriate services in the future.

The performance of a deep learning model significantly improves on challenging datasets when using transfer learning. However, the pre-trained networks have certain constraints in terms of their architecture. For example, the available pre-trained models are trained for a specific input size. Therefore, require resizing the input images of different sizes. When training a model from scratch, higher resolution image offers better performance. However, our study has shown that this is not true when using pre-trained models. We have compared the pre-trained MobileNetV2 performance on CIFAR10 and CIFAR100 datasets. The pre-trained weights of MobileNetV2 are available for image resolutions of 92x92, 128x128, 160x160, 192x192 and 224x224. The performance of the model is evaluated in terms of classification accuracy. Our analysis have shown that for image resolution of 160x160, the pre-trained model has achieved better classification accuracy.

The purpose of this study was to develop and evaluate Point Cloud Data (PCD) deep learning models and a rule-based system for segmenting tree structures (stems and crowns) using fixed terrestrial LiDAR data. The dataset comprised 48 Larix Kaemferi trees, which were collected and preprocessed. For the PCD deep learning models, three downsampled datasets consisting of 1024, 4096, and 16384 points were constructed from the original data. The data was divided into training (70%) and validation (30%) sets. Models were built using PointNet and PointNet++ architectures, resulting in a total of 12 tree structure segmentation models for accuracy comparison. The rule-based system was developed using the original data, applying techniques such as verticality checks, cylindrical structure detection, and slice-based circular fitting to detect the stem. It then segmented the stem through repetitive circle fitting and validation processes based on height. The average accuracy of the PCD deep learning tree structure segmentation models was approximately 95%, with the PointNet++ model using 16384 points achieving the highest classification accuracy of about 98%. The rule-based system achieved high classification accuracy of over 99% for both tree species. This study is expected to contribute to precise measurement and efficient management of forest resources by presenting automated methods for tree structure segmentation using AI technology and rule-based approaches. It is anticipated that this research will serve as a foundation for the advancement of forest digitalization, precision forest management technologies, forest structure analysis, and timber production estimation in various fields.

This research aimed to assess the possibility of detecting forest degradation using time-series satellite imagery and three different deep learning-based change detection techniques. The dataset used for the deep learning models was composed of two sets, one based on surface reflectance (SR) spectral information from satellite imagery, combined with Texture Information (GLCM; Gray-Level Co-occurrence Matrix) and terrain information. The deep learning models employed for land cover change detection included image differencing using the Unet semantic segmentation model, multi-encoder Unet model, and multi-encoder Unet++ model. The study found that there was no significant difference in accuracy between the deep learning models for forest degradation detection. Both training and validation accuracies were approximately 89% and 92%, respectively. Among the three deep learning models, the multi-encoder Unet model showed the most efficient analysis time and comparable accuracy. Moreover, models that incorporated both texture and gradient information in addition to spectral information were found to have a higher classification accuracy compared to models that used only spectral information. Overall, the accuracy of forest degradation extraction was outstanding, achieving 98%.

Lossy image compression provides an efficient solution to the exchange and storage of image data for consumer applications. The design of lossy algorithms is based on a principle to discard information that are not perceivable by human visual system (HVS). With the popularity of deep learning models (DL) in computer vision (CV), it is necessary to characterize the loss in image quality with respect to computer vision systems as well. Recent studies have analyzed the image distortions resulted from blur and noise, mainly from an adversarial attack perspective. However, fewer studies have dealt with the lossy nature of the JPEG algorithm. Therefore, the current study presents a quantitative assessment of different types of data loss that occurs due to chroma subsampling, quantization, and rounding functions of the JPEG algorithm. In addition, we have analyzed impact of different interpolation methods that are used for chroma upsampling. The analysis have shown that for compression savings, performing either subsampling or quantization preserved the model accuracy while their combination degraded the accuracy by 6%.

This article is two-fold. The ultimate goal of this article is to provide a big data analysis of 330 reviews of the movie Noryang and to evaluate the Naive Bayes model, the Random Forests model, the DNN model, and the LSTM model in machine learning and deep learning. A point to note is that the name Yi, Sun-shin was the most widely used by viewers, followed by the word movie, and the word general, in that order. A major point of this article is that the name Yi, Sun-shin and the word movie showed up twice as the first keyword. This in turn implies that these keywords are the most noteworthy ones. The sentiment analysis argues that about 75% of viewers think of the film as well-made and that they were highly satisfied with it. In this paper, we used the Naive Bayes model, the Random Forests model, the DNN model, and the LSTM model and made them predict whether each review is positive or negative. The Random Forests model works well for our data, whereas the Naive Bayes model does not. When learning took place 25 times, the DNN model worked well for our data (its accuracy rate is 82.76%). When it comes to the LSTM model, its accuracy did not improve even though learning took place 9 times. Yet, the LSTM model is slightly better than the DNN model with respect to the accuracy rate of test data.

최근 발생하고 있는 기후변화는 전세계적으로 많은 피해를 발생시키고 있다. 한반도의 기존 강우형태는 장시간에 걸친 강우와 발생기간을 예측할 수 있었으나, 기후 변화로 인한 강우 형태의 변화는 기존에 마련한 위기대응방안이 무용한 상황을 만들고 있다. 2022년 8월 집중호우로 인해 서울·경기·강원·충남 등 10개 지자체가 특별재난지역으로 우선 선포되었다. 도시화에 따른 불투수면적의 증가와 건물의 지하 공간의 활용 증가하고 있는 가운데 기후 변화로 인한 집중호우는 도심지 내 하수관로 설계용량을 초과하 는 경우가 빈번해지고 있으며, 범위는 국부적이지만 피해의 규모는 커지는 사례들이 증가하고 있다. 도시침수는 다양한 조건에 의해 침수범위가 결정되며 불시에 일어나는 침수의 형태로 인해 신속한 침수상황 인지 및 위기대응이 필요하다. 이를 위하여 AI기반 영상분석 기술을 통해 전국에 분포되어 있는 CCTV를 활용하여 침수상황을 상시 모니터링 할 수 있는 시스템 구축이 필요하며, 정형화된 IoT기반 실시간 계측센서 데이터와 비정형 데이터인 CCTV영상을 분석하고 연계한 새로운 형태의 도시침수 모니터링 기술이 필요하다.

Recently, climate change has caused a lot of damage worldwide. Existing rainfall patterns on the Korean Peninsula could predict long-term rainfall patterns and periods, but changes in rainfall patterns due to climate change made crisis response measures useless. Due to heavy rain in August 2022, 10 local governments, including Seoul, Gyeonggi, Gangwon, and Chungnam, were declared as special disaster areas for the first time. As the impermeable area increases due to urbanization and the use of the underground space of buildings increases, the design capacity of sewage pipes is often exceeded, and the scope is local but the damage scale is increasing. Urban floods require rapid situational awareness and crisis response. To this end, it is necessary to establish a system that can monitor the flood situation at all times using CCTV distributed across the country through AI-based image analysis technology, and a new type of urban flood monitoring technology that analyzes and links it based on standardized IoT unstructured data such as real-time measurement sensor data and CCTV images is required.

The enactment of Republic Act 11106 establishes Filipino Sign Language (FSL) as the primary mode of communication within the deaf-mute community in the Philippines. However, this legal recognition has highlighted a significant communication gap between the deaf-mute and non-deaf-mute populations, as the latter typically does not understand FSL. This study introduces a mobile application designed to bridge this gap by translating FSL gestures into textual sentences. The application leverages a CNN-BiLSTM deep learning architecture integrated with Mistral 7B, a state-of-the-art Large Language Model (LLM), to recognize continuous multi-sign gestures and translate them into coherent text. To evaluate the system’s effectiveness, two gesture recognition models were compared based on Word Error Rate (WER), calculated using the Levenshtein distance to measure word-level discrepancies. The 1080p30 model with a stride of 5 and a window size of 30 frames achieved a WER of 27.02%, while the 720p60 model achieved, with a stride of 5 and a window size of 60 frames, a WER of 43.37%. The superior performance of the 1080p30 model is attributed to its higher spatial resolution. This research addresses the critical need for accessible communication tools, offering a solution that enhances inclusivity for the Filipino deaf community.

In this paper, we present a comparative study on performance of deep learning models for electroglottography (EGG) and voice conversion to text using transfer learning. In this regard, we deployed range of deep learning models such as ResNet101, MobileNetv2, GoogleNet for text recognition using electroglottography and voice signals correspondingly. Firstly, short-time Fourier transform (STFT) is utilized to generate spectrogram using time-series signals (EGG, Voice). Spectrogram images are resized to fulfill the requirement of pre-trained models (ImageNet-weights). Subsequently, rigorous experiments have been performed with various combinations of EGG, Voice and hybrid (EGG and voice). In addition, we have studied the impact of healthy and pathology signals using SVD dataset. Expectedly, the accuracies of healthy voice signals were significantly higher as compared to pathology signals. We analyzed the performance of each model under two combinations (healthy and mix). ResNet 101 outperforms other models in terms of generalizability as the accuracies were significantly higher in all three scenarios. The highest accuracy of RestNet 101 in the scenario of healthy and mix for voice signal is 98.10 and 88.57 respectively.

본 연구는 거시경제 환경 변화 속에서 국내 총광고비의 장기적 추세와 예측 가능성을 실증적으로 규명하고자 하였다. 이를 위해 1985–2022년 연도별 총광고비 및 명목 GDP 자료 를 활용하여 광고비의 상대적 불변성 가설(Principle of Relative Constancy, PRC)을 재검토하고, 전통적 시계열 모형과 머신러닝·딥러닝 기반 예측 모형의 성능을 체계적으로 비교하였다. 상대 적 불변성 검증을 위해 시기별 동등성 검정(Two One-Sided Test, TOST)과 선형 추세 분석을 수행하였다. 분석 결과, 광고비/GDP 비율은 장기적으로 0.6%–1.1% 범위 내에서 유지되는 경향 을 보였으나, ±0.1% 허용 구간 기준에서 동등성이 기각되었고, 연도 계수는 유의한 음(-)의 값 을 나타내어 광고비 비중이 점진적으로 감소하는 구조적 추세가 존재함을 확인하였다. 예측 성 능 비교에서는 베이지안 벡터자기회귀(BVAR) 모형이 시계열 모형 중 가장 낮은 오차를 기록하 였고, 딥러닝 모형 가운데 LSTM이 전체 모형 중 최저 RMSE를 나타내어 가장 우수한 예측력을 보였다. 또한 향후 5년 예측 결과, 코로나19 이후 감소한 광고비 비중은 1% 미만 수준에서 비교 적 안정적으로 유지될 가능성이 높은 것으로 나타났다. 본 연구는 PRC 가설의 통계적 타당성을 재검토함과 동시에, 서로 다른 예측 방법론의 상대적 효용성을 실증적으로 제시함으로써 광고산 업의 중장기 전략 수립과 정책적 의사결정에 기여할 수 있는 분석 틀을 제공한다.

This empirically investigates the long-term trend and predictability of total advertising expenditure in Korea under changing macroeconomic conditions. Using annual data on total advertising expenditure and nominal GDP from 1985 to 2022, the study reexamines the Principle of Relative Constancy (PRC) and systematically compares the performance of traditional time series models with machine learning and deep learning-based forecasting models. To test relative constancy, the Two One-Sided Test (TOST) procedure and linear trend analysis were conducted. The results indicate that the advertising expenditure-to-GDP ratio has generally remained within the range of 0.6%-1.1% over the long term. However, equivalence was rejected under a ±0.1% tolerance margin, and the year coefficient showed a statistically significant negative value, suggesting the existence of a gradual structural decline in the share of advertising expenditure. In the forecasting performance comparison, the Bayesian Vector Autoregression (BVAR) model achieved the lowest error among the time series models, while the Long Short-Term Memory (LSTM) model recorded the lowest RMSE among all models, demonstrating the strongest predictive performance. Furthermore, five-year-ahead forecasts suggest that the post-COVID-19 decline in the advertising expenditure share is likely to remain relatively stable at a level below 1%. By reassessing the statistical validity of the PRC hypothesis and empirically comparing the relative effectiveness of different forecasting methodologies, this study provides an analytical framework that can contribute to mid- and long-term strategic planning and policy decision-making in the advertising industry.