Semantic Knowledge Representations for Soft Data Fusion - Efficient Decision Support Systems: Practice and Challenges from Current to Future - page 8

Requests can then be processed on the set of observations, in order to determine whether a
specific situation occurred or not. The study focuses on how to analyze the observation graph
obtained after the fusion of the different information items. As for most of the studies dealing
with graph homomorphism, the authors emphasize on the complexity of the underlying
algorithms.
The authors propose an inexact graph matching technique using a similarity measure between
the nodes and arcs. A model of a situation of interest is drawn, using a graph structure. The
graph matching process then allows finding out whether this model is a sub graph of the
observation graph.
3.3 Defining a high-level information fusion approach and framework
As said before, we focus on information that contains a higher-level of semantics. The
approach that we proposed is optimized for information that has a high level of abstraction
and that is structured. Regarding the JDL model, our approach is suitable for information
fusion of levels 1 and 2 (Object Refinement and Impact Refinement). Level 2 - Comprehension
- of Endsley's model for situation awareness corresponds well to our objectives as well:
synthesis of perceived information items, determination of their relevance to the global
objective of the user and (sub-)situation recognition through the matching with a sough-after
situation.
We propose to use graphs representation formalism, and operations on graph structures.
Representing information thanks to graph structures will allow us first to use existing
operations and theoretical results on graphs. It will also enable to take the advantage of
existing results regarding the optimization of graph algorithms. Our approach is close to
the one proposed in Sambhoos et al. (2008). The aim is to fuse graphs.
However, we do not focus on the algorithmic issues of the problem, but on an other aspect:
solving the conflicts that may appear in the data during the fusion process. When studying
real soft data provided from operational systems, we observed that the different pieces of
information, often contain conflicts. Indeed, as humans are providing the initial input data,
there are very often typing mistakes or differences in the ways the same thing is reported. A
simple example, is when one wants to refer to a person, a first human observer may use the
person's name only, while another one will use name and surname. Therefore, the detection
of conflicts among pieces of information, as well as the resolution of these conflicts within soft
data fusion is of major importance.
We define hereafter the different stages that are necessary to achieve soft data fusion (Figure
4).
Situation & domain modeling is depicted by (1) and (2) on Figure 4. The situation modeling
phase aims at providing a formal definition of the application domain as well as of the
situations that are of interest for the user. The situations of interest are defined thanks to
an empty pattern that describes their characteristics. The objective of the fusion system is
to fill as much as possible this empty pattern with the observations acquired through the
different sources of information.
Soft observations association is depicted by (3) on Figure 4. Observations coming from
different information sources may relate to different objects. They may also relate to the
same object but be reported with small differences. Therefore, it is necessary to determine
whether two incoming observations rely to the same object before to fuse them.