Data mining in bioinformatics is confronted with complex data, which poses challenges w.r.t.
scalability and structured data types. One data structure that is often encountered in this context is
hierarchical data. For instance, gene functions can be structured in a hierarchy from more general
functions (e.g., “receptor activity”) to more specific functions (e.g., “G-protein-coupled
photoreceptor activity”). In many learning tasks, such hierarchically structured background
information is completely ignored.
In this research proposal, we will handle this deficiency in three ways. First, we will exploit the
hierarchical structure of cell types in classifying a patient as having a disease or not, based on the
cells present in his blood sample. Second, we will exploit hierarchically structured cell types and
gene functions in analyzing micro-array data sets. Third, we will investigate the existence of
multiple optimal feature sets in the context of a hierarchically structured target attribute.