Data exploration with self-organizing maps in environmental informatics and bioinformatics

Abstract

The aim of this thesis was to evaluate the usability of self-organizing maps and some other methods of computational intelligence in analysing and modelling problems of environmental informatics and bioinformatics. The concepts of environmental informatics, bioinformatics, computational intelligence and data mining are first defined. There follows an introduction to the data processing chain of knowledge discovery and the methods used in this thesis, namely linear regression, self-organizing maps (SOM), Sammon's mapping, U-matrix representation, fuzzy logic, c-means and fuzzy c-means clustering, multi-layer perceptron (MLP), and regularization and Bayesian techniques. The challenges posed by environmental processes and bioprocesses are then identified, including missing data problems, complex lagged dependencies among variables, non-linear chaotic dynamics, ill-defined inverse problems, and large search space in optimization tasks.

The works included in this thesis are then evaluated and discussed. The results show that the combination of SOM and Sammon's mapping has great potential in data exploration, and can be used to reveal important features of the measurement techniques (e.g. separability of compounds), reveal new information about already studied phenomena, speed up research work, act as a hypothesis generator for traditional research, and supply clear and intuitive visualization of the environmental phenomenon studied. The results of regression studies show, as expected, that the MLP network yields better estimates in predicting future values of airborne pollutant concentration of NO2 compared with SOM based regression or the least squares approach using periodic components. Additionally, the use of local MLP models is shown to be slightly better for estimating future values of episodes compared with one MLP model only. However, it can be concluded in general that the architectural issues tested are not able to solve solely model performance problems.

Finally, recommendations for future work are laid out. Firstly, the data exploration solution should be enhanced with methods from signal processing to enable the handling of measurements with different time scale and lagged multivariate time-series. The main suggestion, however, is to create an integrated environment for testing different hybrid schemes of computational intelligence for better time-series forecasting in environmental informatics and bioinformatics.