4th EucheMs chemistry congress

thursday, 30-Aug 2012
s614
chem. Listy 106, s587–s1425 (2012)
Analytical chemistry Electrochemistry, Analysis, sample manipulation
Chemometrics – i
o - 4 1 9
diSSiMiLArity BASed ModeLinG of CheMiCAL
dAtA
B. wALCzAK 1
1 University of Silesia, Institute of Chemistry, Katowice, Poland
Studying complex non-linear chemical and/or bio-chemical
systems, we need fast and effective methods of their modeling.
As it will be demonstrated, dissimilarity representation (known
also as pairwise representation) reveals data structure, which is
not revealed in the vectorial representation, and thus widens
the sets of possible approaches to data modeling. Performance
of the dissimilarity based methods (e.g., Dissimilarity – Partial
Least Squares ) [1] , will be demonstrated for calibration and
discrimination of real and simulated data sets of different structure
and complexity. Practical aspects of dissimilarity based modeling
(e.g., the choice of dissimilarity measure, pre-selection of the
so-called ‘prototypes’, and fusion of different dissimilarity
measures or different data blocks) will be discussed, as well.
It will also be demonstrated how the dissimilarity based methods
applied to instrumental signals such as, e.g., LC-MS signals, allow
applications of no-warping strategies for data modeling.
As main advantages of the proposed approach, we can
mention its conceptual simplicity, flexibility, and very short
computations time.
references:
1. P. Zerzucha, M. Daszykowski, B. Walczak, Dissimilarity
partial least squares applied to non-linear modeling
problems, Chemometrics and Intelligent Laboratory
Systems, 110 (2012) 156-162
Keywords: non-linear modeling; analysis of variance;
Euclidean distance;
Chemometrics – i
4 th EucheMs chemistry congress
o - 4 2 0
APPLiCAtion of nAture-inSPired MethodS in
CheMoMetriCS
f. MArini 1 , B. wALCzAK 2
1 University of Rome “La Sapienza”, Chemistry, Rome, Italy
2 University of Silesia, Analytical Chemistry, Katowice, Poland
Natural Computing models, inspired in part by nature and
natural systems, are a family of powerful data analysis methods
able to transform available heterogeneous data into knowledge.
They include Neural Networks mimicking the mechanisms of the
nervous system, [1] general optimization techniques, such as
Genetic Algorithms based on simulation of biological evolution [2] ,
Swarm Intelligence based on simulation of social behavior of
animals. [3]
In recent years, many Nature-inspired models have been
successfully applied to the solution of complex problems related
to signal processing, classification, clustering, feature selection,
and regression.
In this communication, some successful application of
natural computing to solve chemical problems will be presented
and discussed. In particular, the use of feed-forward artificial
neural networks to operate non linear classification in the cases
where sample distribution would require complex decision
boundaries will be shown. Additionally, the use of Kohonen
architecture for nonlinear projection aimed at exploratory data
analysis will be exemplified.
As far as Genetic Algorithms are concerned, examples of
their fruitful use in variable selection will be given.
Lastly, some recent applications of a swarm intelligence
algorithm (Particle Swarm Optimization) in chemometrics will be
described. In particular, the use of PSO for parametric time
warping, peak deconvolution and clustering will be discussed.
references:
1. F. Marini, Anal. Chim. Acta 635 (2009) 121-131.
2. R. Leardi, J. Chemometr. 15 (2001) 559-569.
3. J. Kennedy, R.C. Eberhardt, Y. Shi, Swarm Intelligence,
Morgan Kaufmann, 2001.
Keywords: Analytical Methods; Chemoinformatics;
AUGUst 26–30, 2012, PrAGUE, cZEcH rEPUbLIc