키보드 입력 데이터와 심층학습 앙상블 모델을 이용한 파킨슨병 환자 분류

홍연수 경북대학교 데이터사이언스대학원 2024 국내석사

Parkinson's disease is a neurodegenerative disorder characterized by motor symptoms such as bradykinesia, tremor, and rigidity, as well as non-motor symptoms like constipation and depression. However, there is a high rate of misdiagnosis among clinical experts, and significant financial and temporal costs are incurred before receiving a definitive diagnosis at medical institutions. Consequently, machine learning has been increasingly utilized for Parkinson's disease diagnosis since its emergence. Various data types, such as voice signals, handwriting data, EEG signals, and medical imaging data, have been employed to ensure accurate diagnosis given the presence of similar symptoms in numerous conditions. Traditional sensor data requires domain knowledge for feature extraction and selection, often necessitating specialized knowledge in physics and statistics, as well as complex algorithms. Improper feature extraction can hinder the performance of machine learning models. Therefore, this study utilizes keyboard input data, a form of hand motor data that is advantageous for conducting preliminary diagnoses without constraints on time and place for Parkinson's patients. The data is processed using statistical concepts like moments for feature extraction and selection. Moreover, a novel deep learning model named XGB-PDNet is proposed. Despite utilizing relatively small structured data, the proposed approach achieves performance comparable to, or even better in certain classifications, than machine learning techniques that require extensive preprocessing.

지식 그래프 기반 교차 도메인 추천시스템: 음악과 도서간 연결을 중심으로

이우진 경북대학교 데이터사이언스대학원 2025 국내석사

전이학습 기반 Mel-Spectrogram 음성 분석을 통한 파킨슨병 조기 분류

박정훈 경북대학교 데이터사이언스대학원 2025 국내석사

Parkinson’s disease(PD) is a progressive neurodegenerative disorder that significantly affects motor function and speech. Recent advances in voice analysis have shown promise in aiding early diagnosis by identifying subtle vocal impairments associated with PD and related diseases. This study explores the potential of machine learning techniq es to classify and diagnose PD using voice data collected from patients with Parkinson’s disease and other related neurological disorders. By extracting key acoustic features and applying advanced classification models, we aim to improve the accuracy of PD detection and distinguish it from other conditions with overlapping symptoms. The results demonstrate that voice-based analysis can serve as a non-invasive, cost-effective tool to support clinical diagnosis and monitoring of Parkinson’s disease progression.

고속철도 경쟁체제가 기업 재무건전성에 미친 영향 : SRT 개통을 중심으로 한 이중차분 분석

김수정 경북대학교 데이터사이언스대학원 2025 국내석사

본 연구는 SRT 개통으로 도입된 고속철도 경쟁체제가 기업의 재무건전성에 미친 인과적 영향을 분석하였다. 부실기업 수와 지역 간 격차에 초점을 맞추어 재무적 취약성 개선 효과를 규명하였다. 기존 문헌에서는 고속철도 인프라 확장의 경제적 편익이 폭넓게 논의되었으나, 기업 재무건전성에 미치는 영향에 대한 실증적 연구는 제한적이었다. 이에 이중차분법, 강건성 분석, 시간 및 지역 고정효과, 산업별 이질성 분석을 통해 실증적 근거를 제시한다. 분석 결과, SRT와 KTX가 병행 운영되는 실험군 지역에서 연평균 부실기업 발생 빈도가 약 2개 감소하는 정책효과가 확인되었다(DID 계수 –1.977, p=0.001). 그리고 표본에 영향을 크게 미치는 처리집단 지역인 울산광역시, 충청북도를 제외한 강건성 분석과 시간 및 지역 고정효과 모형 적용으로 시간적 외생효과와 지역의 특성을 통제해 결과의 신뢰성과 내적 타당성을 확보하였다. 또한 산업별 이질성 분석에서는 제조업과 도매 및 소매업종에서 특히 뚜렷한 개선 효과를 보였다. 결론적으로 교통 인프라의 경쟁체제 도입이 단순한 지역 접근성 향상을 넘어 기업 재무건전성 개선에 실질적으로 기여하여 기업의 부실화 완화에 영향을 미쳤음을 입증했다. 이러한 결과는 한국의 고속철도 시장에서 공공 독점 구조였던 KTX 중심 체제에 민간 자본이 투입되면서 발생한 구조적 효과에 대한 이해를 심화시키며, 지역 및 산업별 특성을 고려한 교통 인프라 정책 수립의 과학적 근거를 제공한다. This study examines the causal impact of Korea’s high-speed rail competition system, introduced through the launch of the Super Rapid Train (SRT), on corporate financial soundness. Focusing on the number of distressed firms and regional disparities, we provide empirical evidence using a difference-in-differences (DID) approach, robustness checks, fixed effects models, and industry-level heterogeneity analysis. Empirical findings indicate that in regions where both SRT and KTX operate, the number of financially distressed firms decreased by approximately two per year (DID coefficient = –1.977, p = 0.001). Robustness is confirmed through exclusion of influential regions such as Ulsan and North Chungcheong Province and by controlling for time and regional fixed effects. Improvement effects were particularly salient in the manufacturing and wholesale & retail sectors. These outcomes suggest that the introduction of a competitive rail system not only improved physical accessibility but also strengthened financial resilience among firms. By revealing structural effects from transitioning away from a public monopoly to a market with private sector participation, the study provides practical insights for future infrastructure policy that considers regional and industry-specific contexts.

FVTMixer: A Conditional Mixer Framework for Fishing Vessel Trajectory Prediction

이다인 경북대학교 데이터사이언스대학원 2025 국내석사

연속 공정에서 제품 품질 향상을 위한 머신러닝과 베이지안 모델 기반 설계 인자 최적화 연구

조정윤 경북대학교 데이터사이언스대학원 2025 국내석사

Design parameters in manufacturing processes have a direct impact on product quality. Recently, manufacturing companies have been adopting machine learning and deep learning technologies to build smart manufacturing environments and enhance competitiveness. However, research on the optimization of design parameters in continuous manufacturing processes remains limited. This study proposes an integrated prediction–optimization framework that jointly considers time-series process variables and cross-sectional design parameters to improve product quality in continuous manufacturing processes. The framework builds a machine learning–based quality prediction model and applies Bayesian optimization to efficiently identify optimal design parameter settings. Experimental results show that the proposed framework improves product quality by an average of approximately 67% compared to the baseline. In addition, the variance—an indicator of quality fluctuation—was reduced by about 53%, suggesting a more stable and consistent level of quality. These findings demonstrate the effectiveness of the proposed prediction–optimization framework for design parameter optimization in continuous processes and indicate its potential as a practical guideline for achieving both quality improvement and reduction in quality variability in real-world manufacturing environments.

공간 네트워크와 시계열 데이터의 통합 및 분석

이가현 경북대학교 데이터사이언스대학원 2025 국내석사

This study proposes a novel methodology to address missing data at unobserved locations and times in spatial networks. Spatial networks are essential analytical tools in environmental monitoring, disaster management, and urban planning; however, incomplete data collection frequently limits analysis and forecasting accuracy in these applications. To solve this problem, we propose a spatial network and time series data integration algorithm and a space-time coupled interpolation method in spatial networks. The proposed integration algorithm minimizes structural distortion and preserves network functionality when incorporating time-series observation points into the network. Furthermore, by combining a spatial interpolation method that reflects network spatial structure with a newly proposed temporal interpolation method that considers data periodicity and temporal proximity, this approach aims to overcome the limitations of traditional Euclidean spatial interpolation methods and enhance accuracy. Experimental results indicate that the proposed methodology outperforms conventional interpolation techniques in spatial networks, especially demonstrating computational efficiency and cost-saving benefits compared to spatiotemporal kriging. This methodology shows promise for real-time monitoring and predictive models in climate change, environmental pollution, and disaster management, supporting reliable, data-driven decision-making.

물 사용량 데이터를 활용한 인공지능 기반 가정 내 누수 판별

손상현 경북대학교 데이터사이언스대학원 2024 국내석사

As global water management challenges intensify, the early detection of household water leaks has become critical due to its financial and environmental implications. This study proposes a novel AI-based model for detecting leaks using daily water consumption data collected through IoT-enabled remote metering systems, coupled with complaint records and customer information. By systematically labeling leaks and applying a three-day sliding window to create time-series data, the dataset was optimized for machine learning analysis. Various models were evaluated, with hyper-parameter tuning - 47 performed using Random Search, and the proposed model demonstrated superior performance, achieving an F1-score of 86.6% and an Accuracy of 87.8%. It successfully predicts leaks up to two months in advance, offering a proactive approach to mitigate water wastage and prevent financial burdens on households. Beyond its immediate benefits, the methodology is scalable to other regions and adaptable for integration into municipal water management systems. By supporting household financial stability and aligning with global sustainability goals, this research lays the groundwork for innovative, IoT-driven solutions in water resource management. Future directions include integrating environmental data, such as weather patterns, to further enhance model accuracy and applicability.

LISA 공간분석과 스캔 통계를 통한 산불 전후 소나무재선충병의 확산 패턴 변화 및 위험도 평가

이태훈 경북대학교 데이터사이언스대학원 2024 국내석사

Large-scale wildfires can significantly alter forest ecosystems, potentially accelerating the spread of Pine Wilt Disease (PWD). This study investigates how the February 2022 Hapcheon-Goryeong wildfire in Gyeongsangbuk-do, South Korea, influenced the spatial distribution and risk of PWD in Goryeong County. Publicly available data on PWD occurrence, satellite-derived wildfire-affected areas, and administrative boundaries were integrated to perform spatial analyses. First, we employed Kernel Density Estimation (KDE) and Density Ratios to visualize and quantify changes in the spatiotemporal distribution of PWD infection. We then used LISA (Local Indicators of Spatial Association) and SaTScan-based scan statistics to identify and verify significant infection clusters. Additionally, we calculated the Indirectly Standardized Rate Ratio (ISRR) at the local administrative level to determine which areas showed notably higher infection risk. The results revealed that, after the wildfire, there was a marked shift in infection clusters toward the southwestern part of Goryeong County, which had not previously been regarded as a major outbreak zone. In particular, the directly affected area of Ssangnim-myeon (Sinchon-ri) exhibited the highest ISRR (1.87), indicating a statistically significant elevation in infection risk compared to other regions. These findings suggest that trees weakened by wildfires become suitable habitats for the pine sawyer beetle (the primary vector of pine wood nematodes), thereby facilitating the further spread of PWD. By quantitatively confirming that large-scale wildfires are a critical driver of PWD expansion, this study underscores the need for proactive management and enhanced surveillance, particularly in fire-prone areas. Early detection, vaccination, and swift removal of infested trees are recommended strategies to suppress the spread of PWD and help maintain forest health.

공간 특성을 반영한 그래프 신경망 기반 교통사고 위험지역 예측 모형 개발 : 대구시 사례를 중심으로

남혜현 경북대학교 데이터사이언스대학원 2025 국내석사