Hyunsook's Preliminary Homepage

App. Files

About me, click menus on your left or Email me If you want to hire me and are looking for relevant documents, click here. statistical/machine learning, data mining particularly on high dimensional massive data, modeling, prediction, decision theory After loosing 20 and a few lbs, I thoroughly understand that keeping it off is more difficult than loosing it. Keep up calorie counters!!! [Fall, 08] a personal blog [Summer, 07] My full time job running a weblog: AstroStat slog Click this link or the image on the left. [Jan. 01, 07] I began working at CfA (Harvard-Smithsonian Center for Astrophysics) as Postdoctoral Fellow. [Ended 12/31/09] Learning everything in a hard way, making paths no one walked before. I've learned various stuffs and it's time to turn the knowledge into new crops and fruits that nurture astronomers and statisticians. Book list for teaching The Universe is written in the language of mathematics without which it is humanly impossible to understand a single word of it; without this language, we wonder about in a dark labyrinth. - Galileo Galilei

Current Projects

• Ill-posed/Inverse problems in Chandra X-ray Data Analysis
1. ARF: 1dim random curves in an inverse problem (calibration for X-ray spectral fitting);
   estimating not so regular curves nonparametrically and estimated uncertainty of ARFs is ready to be incorporated into spectral fitting
   so that inference results reflect calibration uncertainty properly.
2. RMF: 2dim random matrices/images in inverse problem (calibration for X-ray spectral fitting);
   a RMF indicates the conversion from detector channel to energy channel. Except that it is 2 dimensional, different from 1-dim ARFs, the object,
   i.e. that estimated uncertainty of RMFs is ready to be incorporated into spectral fitting so that inference results (best fits
   and error bars would be familiar terms for astronomers) calibration uncertainty properly similar to ARF uncertainty incorporation.
   Estimating and summarizing random matrices is the challenge.
3. XAtlas: grating spectrum data base, developing tactics for clustering analysis on spectral curves generally very spiky; Poisson count high resolution spectra; not regular spectral curves in statistics; PCA,ICA, logistic regression, nonparametric FDA on processed data.
• Solar Images: ICA for automatic feature detection. A few algorithms are under development
• CMD: get the list of papers that focused on statistical modeling and inference. Recent ones are limited to Bayesian so that with many thousands of stars and with full tabulated isochrone sets, computations become unthinkable. My approaches were taken from information theory. Instead of likelihood being evaluated for thousands and millions times for a Markov chain, a few linear operations for evaluating statistics. By matching information from CMD and from isochorones - although a choice among schools of stellar evolution track computing is opted to be done by the user - the age distribution of a given stellar cluster can be estimated or approximated. My work is distinguished from other Bayesian frame that a priori, instead of feeing all age isochrones, feeding plausible one, on which the likelihood of age distribution is determined. No brute force grid fitting
• LF: luminosity function estimations. Because of truncation and random censoring robost procedure for luminisity function estimation and subsequent inference precedure have not been properly addressed. Not only disussing and formalizing both astronomical and statistical challenges, different attempts are made to solve and to generlize the problem.
• Model Preference Assessment: parameter space wise models of astronomy do not have any commonality but their error distributions are same or Poisson since photons are considered to be subject to Poisson counts. Astronomers do not satisfy with model selection criteria because they do not offer p-values. Partially nested situation, how to modify model selection and measure the degree of model preference?
• log N-log S
• χ²
• Model based clustering data with measurement errors: measurement errors in astronomy are unique and often parametrized into additive error terms in Gaussian distributions. We tested such parameterization helps for better clustering in terms of information or entrophy regardless of addtional parameters violated Occam's razor.
• Source Detection : enhancing and changing current source detection algorithm in ciao
• Nonparametric Contours for 2D: convex hull and Voronoi tessellations for data depth, estimating density including shape measures and outliers.
  Particularly useful displaying bivariate posterior draws which are resprented via heavily populated scatter plots and inefficient in conveying information.
• COSMOS/SDSS: data resources for above and data mining.

Research Interests ---AstroStatistical Problems.

• Nonparametric methods for randomly censored data (incorporating estimated censoring or
  truncation models into the composite model, missing data analysis)
• Modeling and analyzing astrophysical data with measurement errors
• Modeling and analyzing astrophysical data with instrument (systematic) errors
• Calibration uncertainties in Chandra and other instruments
• Feature detection from streaming solar images
• Automatizing Feature Detection
• PCA (objective function: multivariate normal, GLS, ...) and ICA (objective/utility function: hypertangent, information criteria)
• Applying statistical learning algorithms/methodology to various astronomical data and developing/improving them
• Dimension Reduction on Images/Curves for Statistical Inference
• Separating classes in Semi-Supervised Learning Setting
• Automatic Segregation of interesting features from Solar Images
• Statistical approaches to retrieve information from stellar clusters
• Data mining, particularly clustering on astronomical data sets
• Resampling and adaptive filter
• Fitting complex functions and curves from astrophysics
• Automated Morphological Classification of galaxies from surveys
• Marked Point Process for galaxies and remote heavenly objects for revealing the history of the universe
• Fast Nonparametric Estimation for Multi-Dimensional Massive Data
• Robust Model Selection Methods which lead to proper physical interpretation
• Estimating densities through Tessellation (computational geometry)
• (Quasi) Monte Carlo methods and high dimensional data inference

List of Publications

1. A jackknife type approach to statistical model selection with G.J.Babu and C.R.Rao (submitted)
2. H. Lee (2007). Nonparametric Multivariate Analysis of SDSS Quasars by Convex Hull Peeling
   Statistical Challenges in Modern Astronomy IV, Ed. G.J.Babu and E.D.Feigelson, ASP Conf. Ser., 371, 425-6.
3. [Dissertation] Two Topics: A Jackknife Maximum Likelihood Approach to Statistical Model Selection and
   a Convex Hull Peeling Depth Approach to Nonparametric Massive Multivariate Data Analysis with Applications
   Keywords: Kullback-Leibler distance, information criterion, statistical model selection, jackknife, jackknife information criterion (JIC), bias reduction, maximum likelihood, unbiased estimation, non-nested model, computational geometry, convex hull, convex hull peeling, statistical data depth, nonparametric statistics, generalized quantile process, descriptive statistics, skewness, kurtosis, volume functional, convex hull level set, massive data, outlier detection, balloon plot,
4. H. Lee (2006). New Statistical Insights of Globular Cluster Systems
   Astronomical Data Analysis Software and Systems XV, Ed. Gabriel et. al., 351, 181-5
5. H. Lee, H. Yun, and J. Chae (2000). Non-thermal Broadening of UV lines Observed at the Limb of the Quiet Sun
   Journal of the Korean Astronomical Society, 33, 57-73 [Paper]

Papers in Preparation

1. Easy and intuitive but statistical just way to reveal star formation history ---- except a few Bayesian methods,
   relying on complex computation and parametric models, most methods are variations of chi square eye balling
2. Application of Statistical Resampling Methods to the Adaptive Filter Design
3. Improving χ^2 methods in Astrophysics
4. Multivariate Outlier Detection with Convex Hull Peeling
5. Multivariate Descriptive Statistics with Convex Hull Peeling
6. Spatiotemporal Model for Mapping the Universe (z as time component and b and l as spatial components)
7. Tessellation and nonparametric density estimation
8. Information criteria under the unspecified true model
9. Nonparametric Statistical Inference of Multivariate Uniform Distribution with Convex Hull Peeling
   (check Cabo and Groeneboom, 1994; Schneider, 1987)
10. Non Distance Metric Based Clustering Algorithm without Iteration

Miscellaneous

Places I've been to

• Since 01/01/10, Daejon
• Born and raised in Daejon, went SNU and PSU
• 2007-09: lived in Somerville/worked in Cambridge, MA
• May. 2009: Erie, PA (to be a bridemaid, groom and some groomsmen are friends)
• Feb. 2009: Paradise Island, the Bahamas
• Jan. 2009: Gainseville, Fl (drove without break to Key West)
• Aug. 2008: Denver, CO
• June 2008: St. Louis, MO
• Nov. 2007: Shang-hai, Hang-zhu, and So-zhu, China
• Aug. 2007: Salt Lake City, Ut
• Aug. 2006: Seattle, WA
• Spring 2006: Durham, Raleigh (part), Chapel Hill (UNC), NC
• Oct. 2005: Touristic cities in Spain
• Sept. 2005: Denver, Dillon, CO
• Aug. 2005: Minneapolis and St. Paul, MN
• Dec. 2004: Kansas City, MI
• Aug. 2004: Toronto, Canada
• Jan. 2004: Vail, CO
• May. 2002: Chicago, IL
• March 2002/Dec. 2001: San Francisco, CA
• Aug. 2001: Ottawa, Montreal, Boston (with Hyuna)
• Summer 2001: New Brunswick (next to NJ transit station), Middletown (AT&T Labs. 201 Laurel Ave. 07748), NJ
• DC and NYC (relatively close to State College, oTL)
• Sept-Oct. 1998: Big Bear City and LA, CA
• Few cities (Berlin, Rome, Paris, etc) in Europe (Jan, 1996)
• Some cities in Europe (Feb. 1986)
• Lived in Montreal (Canada, 1985-1986) and kept the shock from the Canadian Parliament building in Ottawa for 16 years.