Office: EA 723B
Phone:
(215) 204-3160
Fax: (215) 204-5960
E-mail:
ljpark at temple dot edu
I am an associate professor in Department of
Electronics Engineering, Kangnung National University (KNU) in
Research interests
and specialties
■
Machine learning
■
Pattern recognition
■
Evolutionary computation
■
Sensor networks
Current
research topics
■
Classification in
imbalanced data set, ROC, AUC, and PAUC
Motivation
Design and learning of a classifier for
real-world two-class classification problems are often plagued by highly
imbalanced, severely overlapping class distribution. Examples are database
marketing, fraud detection, and medical diagnosis. In those classification
problems, receiver operating characteristic (ROC) curve has
been used to help us to visualize a
trade-off of classifier's discrimination capability that is indistinguishable
in the traditional accuracy measure. AUC (the area under ROC curve) has been
widely used as a single performance measure to evaluate an average classifier's
performance especially when information on class ratio and/or misclassification
costs is unknown.
Recently, much attention has been paid to learn
classifiers by maximizing AUC in machine learning and data mining communities.
In some classification applications, however, it is often desirable to optimize
classifier's discrimination performance at a certain operating range, not in
the entire operating range as in the AUC. For example, in medical diagnosis,
true positive rates of less than, say, 0.7-0.8 would be probably unacceptable.
Objectives
In order to produce a
high-quality classifier in some real-world applications such as fraud
detection, a method is required that is capable of optimizing classifier's
discrimination performance at a desired local operating range, for example, the
TPR at a certain range of FPRs.
Figure 1.
Partial AUC and decision boundaries on the feature space in fraud
detection.
■
Terahertz signal
restoration
Motivation
There has been a significant interest in adopting
terahertz (THz) signal
technology, spectroscopy, and imaging for security applications.
Without health-risk for scanning of people, the THz radiation is capable of
detecting concealed weapons and illicit drugs because they have characteristic
THz spectra.
Stand-off detection of concealed threat materials in open areas such as
airports and stations requires THz signal transmission in the open air.
Unfortunately, THz pulses through the humid atmosphere are quietly distorted by
atmospheric absorption which is mainly due to polarity of water-vapor. The
distortion is generated by absorption and scattering of water-vapor in the
atmosphere during THz signal propagation through the air from the source to the
spectrometer. In
THz transmission in the open humid air with numerous specific absorption
frequencies makes it difficult to extract THz spectral signatures of concealed
threat materials.
Objectives
A practical stand-off THz sensing system, which enables us to identify
some explosives 20~30m away, should be able to restore the absorptive losses of
THz radiation by the atmosphere. To achieve this, effective signal processing
techniques are required for restoration of a THz degraded pulse in the open
humid atmosphere.
Figure 2. A degraded THz pulse due to the
water-vapor absorption
(Time- and frequency-domain).