http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
연합 학습을 사용한 사용자 건강 정보 예측에 관한 연구
방준일(Bang Junil),홍성은(Hong Sungen),김선욱(Kim Seon Uk),김화종(Kim Hwajong) 한국통신학회 2021 한국통신학회 학술대회논문집 Vol.2021 No.2
본 논문은 Federated Learning을 활용하여 생체정보 및 의료정보를 학습하여 사용자에게 맞춤형으로 건강상태를 예측하고 경고해 주는 전체 프로세스를 설계한 연구이다. 해당 프로세스의 검증을 위한 사전연구 정보와 관련 정보를 서술하였고, 전체 프로세스의 설계 방안을 연구하였으며, 향후 실제 진단 결과를 학습에 반영하여 성능을 고도화하는 방법을 연구하였다.
관절 데이터 기반 동작 인식 모델 연합학습 프레임워크 연구
방준일(Junil Bang),홍성은(Seongeun Hong),전석환(Sukhwan Jeon),이주원(Joowon Lee),김화종(Hwajong Kim) 한국정보기술학회 2023 한국정보기술학회논문지 Vol.21 No.3
This study corresponds to the implementation of federated learning among the systems that help caregivers taking care of many patients in a nursing hospital by photographing a nursing hospital patient with a bedside imaging device and building a motion recognition model with this image. De-identified and lightweight ETRI-Activity3D joint data was used for federated learning of the graph-based motion recognition deep learning model, and lightweight STGCN(Spatio-Temporal Graph Convolutional Networks) based motion recognition model was used for federated learning of time-series graphs. model was modified. Federated learning was implemented based on the open source Flower. The global model collected by the aggregation algorithm in the federated learning client showed better Accuracy than the model using only locally owned original data. Compared to the centralized model performed with the same physical and temporal resources, about 98% of performance was achieved.
IoT 환경에서 GDPR에 부합하는 개인정보수집 동의 절차
이구연(Goo Yeon Lee),방준일(Junil Bang),차경진(Kyung Jin Cha),김화종(Hwa Jong Kim) 한국정보기술학회 2019 한국정보기술학회논문지 Vol.17 No.5
Many IoT devices like sensors lack screen and input devices, thus making them hard to meet the consent conditions that GDPR requires. This is acting as a legal barrier for further advancement in the business field. In this paper, we designed the process for consent of personal information collection that meets the legal conditions. In this design, user’s personal data is received in an encrypted form by data collecting server first. The encrypted personal data can be decrypted after associating with user agent based on the consent procedure of the collection of personal information. During the consent procedure, user agent understands the privacy policy about personal information collection and offers the key to decrypt the data. This kind of personal information collection agreement procedure will satisfy the transparent and freely given consent requirements of GDPR. Thus, we can speculate from here that the proposed procedure will contribute to the evolution of IoT business area dealing with personal information.
시간 정보를 활용한 Time Aligned-LSTM 사람 행동 예측 연구
홍성은(Seongeun Hong),방준일(Junil Bang),김용진(Youngjin Kim),김화종(Hwajong Kim) 한국정보기술학회 2022 한국정보기술학회논문지 Vol.20 No.10
Recently, as IoT devices are widely spread, many sensors exist and measure various information. the house occupies a large part of a humans life, and various sensors can be installed, making it easy to collect various information. By analyzing the users current location, device usage information, and time information, the users activity, patterns, and habits can be found, and activity prediction enables various services. Users Behavior Prediction In previous studies, the time the behavior occurred is very important, but this information was not used for model training. In this study, we propose a users behavior prediction model that uses occurrence time information in addition to sensor data in a smart home environment. The accuracy of the proposed model was 1.2~5.7% higher than that of Bi-LSTM as a result of using the model input of occurrence time and evaluating model performance in multiple data sets.
데이터 임베딩을 활용한 사용자 플레이리스트 기반 음악 추천에 관한 연구
이현수(Hyeonsu Lee),홍성은(Seongeun Hong),방준일(Junil Bang),김화종(Hwajong Kim) 한국정보기술학회 2020 한국정보기술학회논문지 Vol.18 No.9
Recently, the online recommendation system, which is attracting attention, analyzes many variables such as user behavior pattern, item characteristics, and additional variables to recommend items that users want. In this paper, we propose a new method to recommend each item through data embedding and clustering using various catalog formats in the Melon music data set. The proposed method of recommending music based on user playlist using data embedding is used for learning by converting information about songs such as tags, genres, detailed genres, and singer names into a sentence form combined a list of words. The comparison performance evaluation and Item2Vec method of the proposed method are performed based on the similarity of embedded songs by embedding songs in multidimensional vector space through SGNS. As a result, the proposed method improved the recommended performance with an average nDCG 0.2996, compared to the average nDCG 0.1850 of Item2Vec.
ConvLSTM을 사용한 토마토 생산량 및 성장량 예측 모델에 관한 연구
홍성은(Seongeun Hong),박태주(Taeju Park),방준일(Junil Bang),김화종(Hwajong Kim) 한국정보기술학회 2020 한국정보기술학회논문지 Vol.18 No.1
The most important technology is the accurate prediction model of the growth and production of smart farms. However, domestic research is largely based on annual and monthly production forecast studies. Predictive model studies using farm unit data are insufficient, and studies for output forecasting (assumption) are being conducted to derive statistical models, not data-based ones. Therefore, the researcher developed a data-based growth and production prediction model using data from smart farm environments. In the study, multi linear regression, random forest and deep learning algorithm (ConvLSTM) were compared, and the ConvLSTM model, which applied deep learning technique, had the highest R² score for individual and average farmers. The R² score for the production forecast model was 0.981, and the R² for the growth forecast was 0.805.
클래스 불균형 문제에 연합학습 적용을 위한 최적화 기법 연구
이현수(Hyeonsu Lee),홍성은(Seongeun Hong),방준일(Junil Bang),김화종(Hwajong Kim) 한국정보기술학회 2021 한국정보기술학회논문지 Vol.19 No.1
Recently, as highly advanced personal identification technology has made it easier to identify individuals, various measures are required to guarantee the rights of information subjects in the information society. Federated learning is a machine learning approach proposed by these needs, a specific approach to educating machine learning algorithms while keeping the data private. In this paper, in order to identify problems that may arise when applying federated learning to the medical industry, which is sensitive to privacy issues, a retinal patient data set, was disproportionately distributed like the environment in which the actual medical institution holds the data. As a result of experiments applying various learning optimization techniques to class imbalance problems that occur here, F1 score 0.96 was achieved in experiments with under sampling and TopkAvg techniques, and the learning time was also shortened.