REPOSITORY > RESULTS

Abstract

Research in many scientific fields such as physics, biology, medicine, economy etc. requires
processing, analysis and understanding of various types of data with the aim of discovering
new or analysing existing research hypotheses. Increase of available data and different data
sources allows for examining some set of interesting entities from different aspects or using
information about different subunits constituting these entities. This enriches the analysis
and leads to the discovery of potentially interesting, previously unknown, associations.
Redescription mining, a field of knowledge discovery studied in this thesis, uses data
obtained from different sources or describing different aspects of entities, logically organised
into disjoint sets of attributes, to re-describe (find different characterizations of) various
subsets of given entities. This unsupervised data analysis task is very interesting since it
aims to output understandable rule-like descriptions (called redescriptions) that are easy to
interpret by the domain experts. By examining the produced redescriptions, researchers
can notice interesting connections between the attributes, grouped in different logically
connected sets, and can isolate subsets of entities that are interesting for future research.
In the field of biology, it may be interesting to describe different genes based on their
occurrence in different life forms, gene functional profile or their pairwise similarity. In the
field of economy, more specifically world trade, one can describe different world countries by
using information about their trading patterns and their socio-demographic properties. In
medicine, relating different types of indicators (such as clinical, biological etc.) may help in
explaining the observed symptoms and potentially provide interesting research directions
that can lead to the more appropriate patient treatment.
Previous work in redescription mining includes a number of different algorithms that
aim to create accurate and statistically significant redescriptions. A tool for redescription
exploration and analysis of redescriptions has also been developed with the aim of increasing
the overall understanding of produced redescriptions. However, all these techniques
focused on producing or analysing individual redescriptions.
This thesis presents developed methods, techniques and tools for redescription mining,
redescription set construction, optimization and exploration that use the information
about redescriptions and their relations to produce optimized, diverse and accurate sets of
redescriptions. These new techniques allow the users to guide redescription set construction
to affect the structure of the produced redescription set, enable the use of ensembles
of redescription mining methods and provide different modes of targeted and contextual
redescription set exploration. The developed techniques have been applied to problems in
economy (country trade), medicine (Alzheimer’s disease), biology (bioclimatic niches) and
social science (co-authorship of scientific papers).