The Data Mining Solutions Center is working on various
research directions in data mining. Two of these directions are
Event-based mining and Relationship-based mining. Three kinds of
Relationship-based Mining problems are being researched:
- Coincident Association
mining involves market basket analysis to mine for association
rules in vector data.
- Localized Spatial Association
mining involves extracting association rules to characterize textures.
- Temporal Association
mining pertains to mining historical trends.
There are three types of event-based mining problems.
|Type 1: Known Phenomenon, Known Algorithm
know exactly what they are looking for and what sets of algorithms
to use. The mining plan is then set up to perform those operations
over the data sets of interest. The Mesoscale
Convective System (MCS) detection using the SSM/I data set is
one such example of known phenomenon/known algorithm mining performed
|Type 2: Known Phenomenon,
This is known as an algorithm development approach
where the end user knows what phenomenon to target but is unsure
of the characteristics of the phenomenon or what sequence of algorithms
to apply. This is iterative in nature and allows end users the flexibility
of fine tuning their algorithm for the event of interest. Cloud
mask detection using GOES-8 data is an example of known event,
|Type 3: Search for Anomalies
This category is called target independent data
mining. The user searches the data sets for transient events with
thresholds. The main idea is to mine for trends depicting anomalies
in the data sets. This approach could be used to check the quality
of the data set by searching for spurious values or just mining
for rare phenomena which would be hard to detect because of the
size of the data set. Hinke et al (1997) describe such an application
of target independent data mining and the impact in data reduction
while retaining all the transient phenomena information.