Earticle

The reconstruction of weather data is essential for various applications such as weather forecasting, climate research, and disaster preparedness. Traditionally, this task required multiple instruments to record different attributes, posing challenges for complete data reconstruction. In this study, we have proposed a simple yet effective approach based on graph embedding and independent features to reconstruct the entire family of weather attributes. Exploiting weather histories from 62 stations across diverse climate regions in Nepal, our method enables the imputation of temperature and humidity data for specific weather stations as well as all stations over a period of time. Rigorous testing and validation demonstrate the effectiveness of our approach, with key evaluation metrics including Mean Squared Error (MSE), Mean Absolute Error (MAE), and coefficient of determination (R2). Our results highlight the model’s proficiency in reconstructing comprehensive weather data, offering a promising avenue for enhancing the reliability of weather-related applications. Also, the use of graph embedding techniques and independent features, our approach provides a robust framework for reconstructing historical weather data, addressing the challenges associated with incomplete or fragmented datasets.

The rise of wireless devices makes interference a key challenge for reliable communication in dense spectrum-sharing networks. This paper proposes a graph neural network (GNN)-based power control algorithm to minimize the worst-case outage probability by using statistical channel state information (CSI), i.e., position information. By representing the network as a fully connected directed graph with node and edge features derived from transceiver positions, the GNN employs message-passing layers to aggregate interference patterns and infer near-optimal transmit powers. Simulation results demonstrate the scalability and generalization capability of the proposed method, confirming its suitability for real-time deployment in large-scale wireless systems.

In recent years, graph neural networks (GNNs) have been extensively used to analyze graph data across various domains because of their powerful capabilities in learning complex graph-structured data. However, recent research has focused on improving the performance of a single GNN with only two or three layers. This is because stacking layers deeply causes the over-smoothing problem of GNNs, which degrades the performance of GNNs significantly. On the other hand, ensemble methods combine individual weak models to obtain better generalization performance. Among them, gradient boosting is a powerful supervised learning algorithm that adds new weak models in the direction of reducing the errors of the previously created weak models. After repeating this process, gradient boosting combines the weak models to produce a strong model with better performance. Until now, most studies on GNNs have focused on improving the performance of a single GNN. In contrast, improving the performance of GNNs using multiple GNNs has not been studied much yet. In this paper, we propose gradient boosted graph neural networks (GBGNN) that combine multiple shallow GNNs with gradient boosting. We use shallow GNNs as weak models and create new weak models using the proposed gradient boosting-based loss function. Our empirical evaluations on three real-world datasets demonstrate that GBGNN performs much better than a single GNN. Specifically, in our experiments using graph convolutional network (GCN) and graph attention network (GAT) as weak models on the Cora dataset, GBGNN achieves performance improvements of 12.3%p and 6.1%p in node classification accuracy compared to a single GCN and a single GAT, respectively.

With changing scenarios in Beyond 5G (B5G) systems, channel coding methods will be faced with stronger requirements for reliability, flexibility, and low latency. Conventional decoding algorithms such as Belief Propagation (BP) are efficient in structured contexts but not sufficiently flexible to respond effectively to dynamic and noisy channel conditions. In this study, we introduces a new Graph Neural Network (GNN)-based solution to decode Low-Density Parity-Check (LDPC) codes by taking advantage of the intrinsic bipartite graph representation of Tanner graphs. The proposed decoder learns a data-driven message-passing algorithm that performs better than BP over a large range of signal-to-noise ratios (SNRs) and block lengths. The experimental results show that the GNN-based decoder yields much lower bit error rates (BER), faster convergence, and better resistance to noise variability than belief propagation, especially at medium to high SNR levels. Thus, our findings highlight the foundation for scalable, adaptive, and intelligent decoding systems in B5G and 6G architectures.

Since the outbreak of the coronavirus disease 2019 (COVID-19) pandemic, a lot of efforts have been made in the field of data science to help combat against this disease. Among them, forecasting the number of cases of infection is a crucial problem to predict the development of the pandemic. Many deep learning-based models can be applied to solve this type of time series problem. In this research, we would like to take a step forward to incorporate spatial data (geography) with time series data to forecast the cases of region-level infection simultaneously. Specifically, we model a single spatio-temporal graph, in which nodes represent the geographic regions, spatial edges represent the distance between each pair of regions, and temporal edges indicate the node features through time. We evaluate this approach in COVID-19 in a Korean dataset, and we show a decrease of approximately 10% in both RMSE and MAE, and a significant boost to the training speed compared to the baseline models. Moreover, the training efficiency allows this approach to be extended for a large-scale spatio-temporal dataset.

As urban networks integrate heterogeneous radio access technologies (RATs), such as Wi-Fi and 5G/B5G, modeling performance becomes challenging due to interference, spatial variability, and propagation conditions. This paper proposes a graph neural network (GNN)-based framework for predicting throughput, delay, and jitter in multi-RAT environments, considering RAT type. The model encodes network topology and channel characteristics using node and edge features, capturing spatial configuration, congestion, and line-of-sight (LoS) versus non-line-of-sight (NLoS) conditions. The results show that GNNs exhibit robustness across station densities and spatial conditions. The message-passing GNN method performs well for throughput and delay, while non-graph methods better estimate jitter.

최근 활발히 연구되는 딥러닝 방법론은 인공지능의 성능을 급속도로 향상시켰고, 이에 따라 다양한 산업 분야에서 딥러닝을 활용한 시스템이 제시되고 있다. 교통 시스템에서는 GNN을 활용한 공간-시간 그래프 모델링이 교통 속도 예측에 효과적인 것으로 밝혀졌지만, 이는 메모 리 병목 현상을 유발하기 때문에 모델이 비효율적으로 학습된다는 단점이 있다. 따라서 본 연 구에서는 그래프 분할 방법을 통해 도로 네트워크를 분할하여 메모리 병목 현상을 완화함과 동시에 우수한 성능을 달성하고자 한다. 제안 방법론을 검증하기 위해 인천시 UTIC 데이터 분석 결과를 바탕으로 Jensen-Shannon divergence를 사용하여 도로 속도 분포의 유사도를 측정 하였다. 그리고 측정된 유사도를 바탕으로 스펙트럴 클러스터링을 수행하여 도로 네트워크를 군집화하였다. 성능 측정 결과, 도로 네트워크가 7개의 네트워크로 분할되었을 때 MAE 기준 5.52km/h의 오차로 비교 모델 대비 가장 우수한 정확도를 보임과 동시에 메모리 병목 현상 또 한 완화되는 것을 확인할 수 있었다.

Deep learning methodology, which has been actively studied in recent years, has improved the performance of artificial intelligence. Accordingly, systems utilizing deep learning have been proposed in various industries. In traffic systems, spatio-temporal graph modeling using GNN was found to be effective in predicting traffic speed. Still, it has a disadvantage that the model is trained inefficiently due to the memory bottleneck. Therefore, in this study, the road network is clustered through the graph clustering algorithm to reduce memory bottlenecks and simultaneously achieve superior performance. In order to verify the proposed method, the similarity of road speed distribution was measured using Jensen-Shannon divergence based on the analysis result of Incheon UTIC data. Then, the road network was clustered by spectrum clustering based on the measured similarity. As a result of the experiments, it was found that when the road network was divided into seven networks, the memory bottleneck was alleviated while recording the best performance compared to the baselines with MAE of 5.52km/h.

Late blight, caused by Phytophthora infestans, threatens tomato and potato crops in Nepal. This study explores developing and deploying a mobile application powered by a graph neural network (GNN) to predict late blight risk for Nepali farmers. The GNN trained on 43 years of NASA satellite weather data can generate 5-days forecast for 320 weather stations in Sudurpashim and Karnali Province, Nepal. The mobile application offers user-friendly forecasts and visualizes late blight risk through clear graphical interfaces. In the visited sites, 30% of tomato and potato crops were found infected with late blight, which the app had identified as high-risk. Samples infected with late blight were collected and analyzed in a wet lab setting. All infected samples tested positive for P. infestans, confirming the app's ability to predict real-world late blight outbreaks. This study showcases the successful development and deployment of a GNN-powered mobile application for assessing late blight risk in Nepal. The application disseminates critical weather information and localized risk assessments, potentially enhancing late blight management in tomato and potato crops. Further research, including extensive field trials comparing with farmers' practices, could increase the application's usability in Nepali fields.

Our research aims to enhance the modeling of speech signals for more effective extraction of node features and analysis of relationships between nodes. To achieve this, we model speech signals as cyclic or linear graphs. Our model combines layers of Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) to leverage their respective strengths in processing graph data. Specifically, we utilize GCN to aggregate information from neighboring nodes, which helps capture local relationships among nodes. Additionally, we employ GAT mechanisms to assign varying attention weights to different neighboring nodes, facilitating a better capture of complex global relationships between nodes. In our experiments, we validate our approach using the IEMOCAP dataset and demonstrate comparable performance to state-ofthe- art models in emotion recognition tasks. This research outcome provides new insights and methodologies for further exploration in the field of speech signal processing.

Recently, a graph neural network has played a crucial role across various fields. In this paper, we designed a Graph Convolutional Network (GCN) to analyze population movement at the city level. It consists of four Graph Convolution (GC) layers, with each layer responsible for aggregating knowledge from its neighboring nodes and updating the feature representation for each city. We utilized population mobility data from China, which includes daily city-to-city movement data. GCN estimates the strength of relationships among all cities. Experimental results demonstrate that the proposed GCN achieves improved performance in estimating city-to-city migration flow relationships.

The rapid growth of e-commerce platforms demands recommendation systems that balance scalability with representational expressiveness. While Matrix Factorization (MF)-based collaborative filtering offers computational efficiency, it fails to capture high-order contextual relationships among items. Graph Neural Networks (GNNs) address this limitation but incur prohibitive computational costs in large-scale, real-time settings. To resolve this tradeoff, we propose a unified hybrid framework integrating MFbased candidate generation and GNN-based re-ranking within a retrieve–then–rank architecture. The MF stage efficiently generates a candidate set, while the GNN stage refines rankings via a heterogeneous graph incorporating user–item interactions, item attributes, and category structures. We further introduce an availability-aware scoring mechanism that integrates real-time stock information to enhance practical applicability. Extensive experiments on a large-scale e-commerce dataset demonstrate that the proposed framework consistently outperforms baseline methods, effectively mitigates data sparsity, and maintains computational efficiency suitable for real-world deployment.

In recent years, the anticipation of human mobility flow has significant applications in various domains ranging from urban planning to public health. This study proposes a hybrid Graph Neural Network and Long Short Term- Memory network-based model for nationwide human mobility prediction, effectively capturing inter-urban movement patterns. We validate the feasibility and effectiveness of our model using the Korean internal-city mobility dataset, which captures real-world population movement patterns across various urban regions. Our experimental results accurately predict inter-city mobility, advancing urban planning, health, and transport.

댐 유입량의 정확한 예측은 홍수 제어 및 수자원 관리의 핵심 요소이다. 본 연구는 시공간 그래프 신경망(ST-GNN)을 활용한 댐 유입량 예측에서 공간 정보의 기여도를 정량적으로 평가하였다. 충주댐(6,648 km²)과 소양강댐(2,703 km²)을 대상으로 ST-GNN과 5종 비교모형(LSTM, GRU, TCN, Transformer, LSTM-Attention)을 5개 선행시간(T+3h~T+48h)에 대해 비교하고, 소거실험을 통해 그래프 합성곱 계층 유무에 따른 ΔNSE를 산정하였다. 충주댐에서는 T+12h부터 ΔNSE가 양(+)으로 전환되어 T+24h에서 +0.028로 최대치를 기록하였으며, 블록 부트스트랩 분석 결과 95% 신뢰구간이 [+0.010, +0.111]로 통계적 유의성이 확인되었다. 반면 소양강댐에서는 전 선행시간에 걸쳐 음(-)의 ΔNSE가 산출되어(T+24h: -0.035), 급경사 단일 주하천 유역에서 그래프 기반 공간 집계가 예측 성능을 저하시키는 것으로 나타났다. 이러한 결과는 그래프 기반 공간 정보의 유효성이 하천망 구조 및 수문 응답 특성에 따라 유역별로 상이함을 실증하며, 댐 유입량 예측에서 유역 특성을 반영한 모형 선택 전략의 정량적 근거를 제시하였다. 다만, 두 유역에 한정된 결과로서 일반화를 위해서는 다양한 수문학적 특성을 가진 유역으로의 확장 연구가 필요하다.

Accurate prediction of dam inflow is essential for flood control and water resources management. This study quantitatively evaluated the contribution of spatial information in dam inflow prediction using Spatio-Temporal Graph Neural Networks (ST-GNN). ST-GNN and five baseline models (LSTM, GRU, TCN, Transformer, LSTM-Attention) were compared for Chungju Dam (6,648 km²) and Soyanggang Dam (2,703 km²) across five lead times (T+3h to T+48h). An ablation study quantified spatial information contribution by computing ΔNSE between ST-GNN with and without graph convolution layers. For Chungju Dam, ΔNSE turned positive from T+12h, reaching +0.028 at T+24h, with block bootstrap analysis confirming statistical significance (95% CI: [+0.010, +0.111]). In contrast, Soyanggang Dam showed negative ΔNSE across all lead times (T+24h: -0.035), indicating that graph-based spatial aggregation degraded performance in a steep, single-channel basin. These results demonstrate that the effectiveness of graph-based spatial information is basin-dependent, governed by stream network complexity and hydrological response characteristics, providing quantitative evidence for basin-specific model selection strategies.

기업 부실 예측 모델은 부도와 같은 사후적 사건에 집중하여, 그 전 단계의 중요한 위험 신호인 비적정 감사의견을 예측하는 데 한계가 있었습니다. 기업 간 상호작용으로 전파되는 네트워크 리스크를 포착하지 못하는 한계가 있었다. 본 연구는 이러한 한계를 극복하고자 그래프 신경망(GNN)을 활용하여 기업의 구조적 연결성을 반영한 비적정 감사의견 예측 모델을 제안한다. 금융감독원 DART 공시 데이터를 기반으로 2,057개 상장기업의 5년간 재무 정보와 감사의견을 수집하고, 기업 간 출자 관계와 감사인 공유 관계를 통합하여 약 45만 건의 엣지로 구성된 이중 관계 네트워크(Dual-edge Network)를 구축하였다. 기업의 잠재적 부실 징후는 외부감사인의 비적정 감사의견을 기준으로 정의하였으며, GCN, GraphSAGE, GATv2 등 다양한 GNN 모델을 적용하여 로지스틱 회귀 및 XGBoost와 같은 전통적 모델과 성능을 비교 평가하였다. 실험 결과, 하이퍼파라미터 튜닝을 거친 GraphSAGE 모델이 비적정 감사의견 기업(소수 클래스) 탐지 성능에서 F1-Score 0.551과 ROC-AUC 0.936을 기록하며 종합적으로 가장 안정적인 예측력을 보였다. 이는 기업 네트워크 내 위험 신호를 효과적으로 일반화하여 학습하는 능력의 중요성을 입증한다. GNNExplainer를 통한 분석은 모델 예측의 해석 가능성을 제시했으며, 본 연구 결과는 금융기관, 투자자, 감사인이 기업의 감사 리스크(Audit Risk)를 사전에 식별하는 조기 경보 시스템으로 활용되어 실무적으로 기여할 수 있다. 향후 시계열 데이터를 반영한 동적 그래프 모델링 및 공급망 등 다양한 관계를 통합하는 연구로 확장될 수 있을 것이다.

Traditional models for predicting corporate failure or audit risk have relied heavily on the financial indicators of individual firms, failing to capture network risks that propagate through inter-firm interactions. To address this limitation, this study proposes an audit risk prediction model that incorporates structural connectivity among companies using Graph Neural Networks (GNNs). Based on public disclosure data from the Financial Supervisory Service’s DART system, we collected five years of financial information and audit opinions for 2,057 listed companies. By integrating equity investment relationships and shared auditor connections, we constructed a dual-edge network consisting of approximately 450,000 edges. Audit risk was defined according to external audit opinions, and various GNN models—including GCN, GraphSAGE, and GATv2—were evaluated and compared with traditional models such as logistic regression and XGBoost. Experimental results show that the hyperparameter-tuned GraphSAGE model achieved the most stable and comprehensive predictive performance in detecting at-risk firms (minority class), with an F1-Score of 0.551 and an ROC-AUC of 0.936. This highlights the importance of effectively learning and generalizing risk signals embedded within corporate networks. Moreover, analysis using GNNExplainer provided interpretability into the model’s predictions. The findings of this study offer practical implications for improving audit risk management systems and advancing network-based risk management in financial institutions. Future research can be extended to dynamic graph modeling by incorporating time-series data and integrating additional relational dimensions such as supply chain links.

본 논문은 음성신호를 순환 또는 선형그래프로 모델링하여 노드 특성 추출 및 효과적인 노드 간 관계 분석을 기반으 로 하는 음성감정인식 방법을 제안한다. 제안한 방법은 그래프 데이터를 처리하는 데 있어 각각의 강점을 활용하기 위해 그래프 합성곱 신경망(GCN)과 그래프 어텐션 신경망(GAT) 계층을 결합한다. 구체적으로, GCN을 사용하여 이웃 노드로부터 정보를 집계함으로써 노드 간의 로컬 관계를 포착한다. 또한, GAT 메커니즘을 활용하여 다양한 이웃 노드에 가중치를 부여하여 노드 간의 복잡한 글로벌 관계를 더 잘 포착할 수 있도록 한다. 실험에서는 IEMOCAP 데이터셋을 사용하여 우리의 접근 방식을 검증하고, 감정 인식 작업에서 최신 모델과 견줄 만한 성능을 보여준다. 이 연구 결과는 음성 신호 처리 분야에서 추가 탐구를 위한 새로운 통찰력과 방법론을 제공한다.

This paper proposes a speech emotion recognition method based on modeling speech signals as circular or linear graphs, enabling the extraction of node characteristics and practical analysis of relationships between nodes. The proposed method combines Graph Convolutional Networks(GCN) and Graph Attention Networks(GAT) layers to leverage the strengths of each in processing graph data. Precisely, GCN captures local relationships between nodes by aggregating information from neighboring nodes. The GAT mechanism better captures complex global relationships between nodes by assigning weights to neighboring nodes. Experiments validate our approach using the IEMOCAP dataset, demonstrating performance comparable to state-of-the-art models in emotion recognition tasks. The results of this study provide new insights and methodologies for further exploration in the field of speech signal processing.

Traffic flow prediction is of great significance in urban planning and traffic management. As the complexity of urban traffic increases, existing prediction methods still face challenges, especially for the fusion of spatiotemporal information and the capture of long-term dependencies. This study aims to use the fusion model of graph neural network to solve the spatio-temporal information fusion problem in traffic flow prediction. We propose a new deep learning model Spatio-Temporal Information Fusion using Graph Neural Networks (STFGNN). We use GCN module, TCN module and LSTM module alternately to carry out spatiotemporal information fusion. GCN and multi-core TCN capture the temporal and spatial dependencies of traffic flow respectively, and LSTM connects multiple fusion modules to carry out spatiotemporal information fusion. In the experimental evaluation of real traffic flow data, STFGNN showed better performance than other models.

본 논문은 그래프 합성곱 신경망을 이용한 신경망 구조 탐색 모델 설계를 제안한다. 딥 러닝은 블랙박스로 학 습이 진행되는 특성으로 인해 설계한 모델이 최적화된 성능을 가지는 구조인지 검증하지 못하는 문제점이 존재한다. 신경망 구조 탐색 모델은 모델을 생성하는 순환 신경망과 생성된 네트워크인 합성곱 신경망으로 구성되어있다. 통상 의 신경망 구조 탐색 모델은 순환신경망 계열을 사용하지만 우리는 본 논문에서 순환신경망 대신 그래프 합성곱 신 경망을 사용하여 합성곱 신경망 모델을 생성하는 GC-NAS를 제안한다. 제안하는 GC-NAS는 Layer Extraction Block을 이용하여 Depth를 탐색하며 Hyper Parameter Prediction Block을 이용하여 Depth 정보를 기반으로 한 spatial, temporal 정보(hyper parameter)를 병렬적으로 탐색합니다. 따라서 Depth 정보를 반영하기 때문에 탐색 영 역이 더 넓으며 Depth 정보와 병렬적 탐색을 진행함으로 모델의 탐색 영역의 목적성이 분명하기 때문에 GC-NAS대 비 이론적 구조에 있어서 우위에 있다고 판단된다. GC-NAS는 그래프 합성곱 신경망 블록 및 그래프 생성 알고리즘 을 통하여 기존 신경망 구조 탐색 모델에서 순환 신경망이 가지는 고차원 시간 축의 문제와 공간적 탐색의 범위 문 제를 해결할 것으로 기대한다. 또한 우리는 본 논문이 제안하는 GC-NAS를 통하여 신경망 구조 탐색에 그래프 합성 곱 신경망을 적용하는 연구가 활발히 이루어질 수 있는 계기가 될 수 있기를 기대한다.

This paper proposes the design of a neural network structure search model using graph convolutional neural networks. Deep learning has a problem of not being able to verify whether the designed model has a structure with optimized performance due to the nature of learning as a black box. The neural network structure search model is composed of a recurrent neural network that creates a model and a convolutional neural network that is the generated network. Conventional neural network structure search models use recurrent neural networks, but in this paper, we propose GC-NAS, which uses graph convolutional neural networks instead of recurrent neural networks to create convolutional neural network models. The proposed GC-NAS uses the Layer Extraction Block to explore depth, and the Hyper Parameter Prediction Block to explore spatial and temporal information (hyper parameters) based on depth information in parallel. Therefore, since the depth information is reflected, the search area is wider, and the purpose of the search area of the model is clear by conducting a parallel search with depth information, so it is judged to be superior in theoretical structure compared to GC-NAS. GC-NAS is expected to solve the problem of the high-dimensional time axis and the range of spatial search of recurrent neural networks in the existing neural network structure search model through the graph convolutional neural network block and graph generation algorithm. In addition, we hope that the GC-NAS proposed in this paper will serve as an opportunity for active research on the application of graph convolutional neural networks to neural network structure search.

The rapid advancement of machine learning (ML) has revolutionized molecular property predictions with achieving remarkable accuracy. However, their black-box nature limits interpretability, making it challenging for chemists to extract scientific insights and validate predictions against established chemical principles. To address this, Shapley Additive Explanations (SHAP) have been widely adopted, yet their application to graph neural networks (GNNs) remains challenging. Here, we develop a modified SHAP strategy to extract atom-wise contribution values from GNN predictions. We apply this approach to GNN models predicting fuel reactivity (cetane number) and Gibbs free energy of solvation. Our method provides chemically meaningful interpretations, aligning SHAP-derived descriptors with known chemical knowledge, including fuel's reactivity and solvation effects. The results demonstrate that atom-wise SHAP explanations offer valuable insights into molecular properties without requiring expensive quantum-mechanical calculations, enhancing the interpretability of ML-driven chemical predictions.

인공지능 기술이 발전함에 따라 인공지능은 인간이 하는 업무들을 대체하고 있다. 현재 인공지능 기술은 시각, 청각 분야로 초점이 맞춰져 있으나 최근 후각 분야에 관련된 연구도 활발히 진행 중이다. 후각 AI는 식품, 의료, 보안, 안전 등에 활용될 전망이며 본 논문에서는 우리 사회의 안전불감증 문제를 언급하고 오작동 비율이 높은 화재경보기에 후각 AI를 대입하여 화재경보기의 오작동 비율을 줄이고 화재경보기에 대한 인식을 해결되는 것을 기대한다.

본 논문은 챕터 수준의 강의 동영상 추천 시스템에 있어서 추천의 정확도와 처리속도 간의 균형문제, 즉, 추천 정확도를 향상시키려면 처리 속도가 저하되고, 반대로 처리 속도를 높일 경우 정확도가 감소하는 문제에 대하여 연구한다. 본 논문에서는 이의 해결을 위하여 TF-IDF, K-Means++ Clustering, Graph Neural Network(GNN) 등 다양한 기법을 복합적으로 활용하는 방법을 제안한다. 즉, 챕터들의 유사성을 바탕으로 클러스터를 사전에 구성함으로써 검색 시의 계산량을 줄여 속도를 향상시키면서도, 클러스터를 노드로 하는 그래프에 대하여 GNN을 적용함으로써 추천의 정확도를 향상시키는 방법을 제안한다. 실험 결과 GNN을 사용한 경우 추천의 정확도가 MRR 지표에서 약 19.7% 증가하였으며, 유사도 기반의 정밀도에 있어서 약 27.7% 증가하는 결과를 확인할 수 있었다. 이를 통해 학습자의 질의에 보다 적합한 동영상 챕터를 추천하는 학습시스템 구축에 기여할 것으로 기대한다.

In this paper, we propose a system for recommending lecture videos at the chapter level, addressing the balance between accuracy and processing speed in chapter-level video recommendations. Specifically, it has been observed that enhancing recommendation accuracy reduces processing speed, while increasing processing speed decreases accuracy. To mitigate this trade-off, a hybrid approach is proposed, utilizing techniques such as TF-IDF, k-means++ clustering, and Graph Neural Networks (GNN). The approach involves pre-constructing clusters based on chapter similarity to reduce computational load during recommendations, thereby improving processing speed, and applying GNN to the graph of clusters as nodes to enhance recommendation accuracy. Experimental results indicate that the use of GNN resulted in an approximate 19.7% increase in recommendation accuracy, as measured by the Mean Reciprocal Rank (MRR) metric, and an approximate 27.7% increase in precision defined by similarities. These findings are expected to contribute to the development of a learning system that recommends more suitable video chapters in response to learners' queries.