112
I.M. Schleiter . M. Obach . R. Wagner' H. Werner' H.-H. Schmidt
. D. Borchardt
mean squared error (MSE), the root mean squared error (RMSE), and the
coefficient of determination (B) (Bishop 1995).
Finer quality measures are provided below.
7.3
Data Exploration with Unsupervised Learning Systems
An initial application of ANNs in the process of data exploration is the use of
unsupervised trained networks to visualise similarity of patterns, grouping of
objects, and again outlier detection, but in multiple dimensions (Chon et al. 1996;
Foody 1999). One of the most famous techniques is Kohonen's self organised
map, SOM (Kohonen 1995).
Ne2
N&4
Z02
Z03
Ne5
Ku4
Ba4
Ba5
Ba6
Ku6
Frti
Fr4
Ma5
Kul
8al
Ne6
Me6
Ne7
Wel
RW3
RW5
Ma7
Ba2
Fr2
RW7
RW2
Re
RWl
RWS
Ku2
Lo2
lo3
la7
LoS
L04
LoS
Fr1
Re'
Figure 7.1. Similarities of the 55 sampIes with regard to eight water properties
(water temperature, oxygen, conductivity, total nitrogen, total phosphorus, hydromorphological quality dass, substrate diversity, number of species per sampling
site) visualised by aU-matrix of a 15x12 SOM (spatial proximity and light grey
shades mean similarity).