Ada Fu (Computer Science and Engineering, CUHK) 

Title: Connectivity-based versus Density-based Outlier Detection

Abstract: Outlier detection is concerned with discovering the exceptional
behaviors of objects given a large amount of data. There are some 
important applications such as credit card fraud detection, discovering
criminal behaviors in e-commerce, discovering computer intrusion, etc.
In this talk, we first outline some known methods in outlier detection, 
with a model for several such schemes, and describe a compatibility 
theory, which establishes a framework for describing the capabilities of 
detection schemes in terms of matching users' intuitions.
In terms of this compatibility, the density-based scheme is more powerful 
than the distance-based scheme when a data set contains patterns with
diverse characteristics. Density-based scheme, however, is less
effective when the patterns are of comparable densities as the outliers. 
We then introduce a connectivity-based scheme that improves the 
effectiveness of the density-based scheme when a pattern itself is of 
similar density as an outlier. We may compare the density-based and
connectivity-based schemes, in terms of their similarities,
dualities, and strengths and weaknesses. Empirical analysis results help 
to demonstrate the applications with different features where different 
methods are more effective than the others.