Interactive Visual Exploration and Analysis - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 156

data attribute, such as membership in a cluster, which can subsequently be used
together with all other data to improve the analysis.
4. Proprietary analysis (level 4): This class is a container for everything beyond
complex analysis and includes, e.g., the integration of application-specific feature
definitions (such as flow feature detectors [ 2 ]) or could entail the integration
of higher-level feature definition languages. Identifying common concepts and
refining IVA beyond this level is a subject of future research.
We note that this terminology is potentially controversial and that “relational analy-
sis” and “complex analysis” have other possible meanings. Consequently, we present
this classification as a starting point that can evolve as research in IVA progresses.
In the following, we describe the higher levels of IVA in greater detail.
15.4 Relational Analysis
Relational analysis takes the selections in form of brushes and provides means to
combine these brushes (or selections) into more complex feature definitions. A sim-
ple feature definition language uses Boolean expressions, for example, to combine
brushes into more complex feature definitions. Figure 15.3 shows an example from
the analysis of three-dimensional gene expression data. Here, positions correspond
to the locations of cells in an organism, and the multiple fields represent expression
values of genes, i.e., they specify whether a certain gene is expressed in a given
cell. Individual brushes select expression patterns based on single genes. Combining
these brushes using Boolean operations, it is possible to define complex selections.
The example in the figure uses this capability, to combine patterns based on a priori
knowledge about how genes interact, and verify whether the genes involved com-
pletely explain the arising pattern.
It is possible to generalize logical operations to smooth brushes [ 5 , 6 ] and enable
F+C visualization in relational analysis. One associated challenge is to extend the
visual means, which discriminate data subsets in focus from their context, in such a
way that takes this more complex form of feature definition into account. Within each
view, an appropriate F+C visualization is necessary to reflect the brush(es) applied
to this view. Another level of F+C visualization must reflect the overall feature
specification, possibly also involving multiple features. One possible solution to this
problem is a four-level F+C visualization approach proposed by Muigg et al. [ 23 ],
which, as one particular aspect, is based on an intelligent color combination scheme.
Combining brushes usually defines a relation between multiple fields. Early work
on query-driven visualization (QDV) [ 29 ] used similar concepts in that it defined fea-
tures as a Boolean combination of relational expressions. However, in this QDV work,
the features and expressions were known a priori and not refined during analysis. An
important aspect of QDV visualization is the use of indices, such as FastBit [ 33 ], to
accelerate data selection based on queries. However, there is also work on combining
QDV concepts with IVA, e.g., using parallel coordinates [ 27 ].