KyungHee Kim
Sungkyunkwan University, Master’s degree, Jan. 2021. - Present. Major in Statistics, Time series.
Currently I am a researcher at Research Institute of Applied Statistics, Sungkyunkwan University(성균관대학교 응용통계연구소).
Duksung Women’s University, Bachelor’s degree, Jan. 2016 - Jan. 2021. Major in Statistics, minor in Computer science.
Paper: Beyond_Value_Perturbation_Local_Differential_Privacy_in_the_Temporal_Setting.
There are 3 types of temporal DP mechanisms, Forward perturbation, Backward perturbation and Threshold Mechanism. Forward perturbation and Backward perturbation mechanisms are baselines for Temporal DP. However, both of them have a lot of costs especially in missing, repetition and empty cost. Wheareas, Threshold Mechanism doesn’t have these costs at all.
Seminar about Random Graph Network. Mainly cover poisson random graph, small world and Scale free network model. These are fundamental mathmatical models in network analysis. Models have been developed to express real world.
Seminar on concepts, motivation and application of differential privacy. Mainly reviewed: Differential privacy: A survey of results (Dwork, 2008).
김경희. 2022. 서울시민카드 사용자 유입과 이탈에 대한 시계열 이상치 탐지와 DTW 클러스터링. 통계연구, 22(0) : 63-76
The study analyzes the inflow and departure of Seoul citizens’ card users by gender and age group. First, fit the ARMA model and get estimates to detect outliers in univariate time series. Secondly, Dynamic Time Warping (DTW) clustering subsequently enables identify bound clusters. This study aims to help develop the use of Seoul citizens’ cards by enhancing the understanding of user behaviours.
Anomaly detection; AD As Industry 4.0 accelerates system automation, consequences of system failures can have significant social impact(Lee 2008). To prevent this failure, the detection of anomalous is more important. However, the paper introduces pitfalls of TAD evaluation and suggests some other methods to evaluate TAD rigorously.
In bayesian framework, mostly local prior is used for model select. In this paper, nonlocal prior would be better to model select for some reasons.
Recent time series data tend to be very high-dimensional and high-frequency.
Due to heavy computation, many studies have been conducted on the dimension reduction method to efficiently handle classification and clustering. (e.g. DFT, DWT, PAA)
In this paper, DUHT(Discrete Unbalanced Haar Wavelet Tramsformation, Fryzlewicz, 2007) is mainly used.
Predicting landslides in Gyeongsang Province.
Tool: R, Python
Method: lgbm, xgboost, random forest.
Suggestions to reduce resource consumption.
Suggestions for movie box offcie success factors.
Tool: R
Method: Factor analysis, K-means clustring, regression.
Suggest ideas and applications related to shared spaces of universities.
Tool: R (Web app is also made by R, Shiny package).