Views: 2 | Downloads: 0
The thesis addresses semantic subgroup discovery (SSD), a task at the intersection of
relational data mining and semantic web technologies. While subgroup discovery involves
finding statistically most interesting population subgroups (e.g., subgroups that are large
and have unusual statistical characteristics with respect to some property of interest), in
SSD we exploit ontological and other structural background knowledge to improve the
subgroup discovery process.
We developed a formal framework for semantic subgroup discovery and illustrated the
approach on a motivating example. We designed and implemented new algorithms, which
we applied and evaluated in several settings. The new approaches are implemented in SSD
algorithms SDM-SEGS, SDM-Aleph, and Hedwig, which were systematically evaluated
using a number of evaluation measures on two microarray datasets. In statistical validation,
Hedwig proved to be the most successful, followed closely by the others, as no approach
dominated in terms of all the evaluation measures. The developed software is open-source
and available as Python packages, as well as widgets in the ClowdFlows data mining
platform.
Our approaches were applied to three real-life applications: explaining subgroups of
breast cancer patients, multi-resolution 0-1 analysis of DNA aberration data, and subgroup
discovery on financial news articles. In the first application we used the Gene Ontology
as background knowledge used in SSD to generate explanations of patient subgroups; the
approach was compared to the supporting factors methodology. Our results agreed with a
previous study of breast cancer grades. Furthermore, the methodology was made available
as a ClowdFlows workflow, making the experiment repeatable and easily adaptable to
similar problems. In the DNA aberration study, SSD was part of a three-part methodology,
together with clustering through mixture modeling and banded matrices used for innovative
cluster and rule visualization. The methodology provided novel insights on DNA aberration
data. Furthermore, the approach was applied on four publicly available datasets, together
with background knowledge collected from DBpedia. The methodology proved useful on
all but one dataset, where the data did not demonstrate any banded structure at all.
In the financial news application, we wanted to gain insight into a vast collection
of news articles; more specifically, we investigated the relationship between the nancial
market perception of a financial entity and the articles mentioning it. We chose Portugal
as the target financial entity, together with a semi-automatically constructed ontology. To
model the market perception we used the credit default swap (CDS) price, reecting the
probability that a country will be unable to repay its debt. Using Hedwig we found two
interesting rules describing the local extremes in the CDS price.
The thesis also contributes significantly to the field of relational data mining (RDM);
we developed a Python library and widgets for the data mining platform ClowdFlows.
The package includes a number of RDM approaches, as well as support for MySQL and
PostgreSQL, all through a simple API. The package aims to alleviate many of the issues a
researcher new to RDM might encounter.