Introduction - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 22

SICAMM). The model consists of a sequential Bayes processor formulated from
HMM theory. The parameters estimated by Mixca (mixture matrices, centroids,
and source probability densities) in conjunction with the class-transition proba-
bilities are used for classification. SICAMM is applied in simulations and real data
analysis. This analyses consists of processing features extracted from 8-hours EEG
signals to estimate a two-class (wake and sleep) hypnogram [ 89 ]. Both simulated
and real data demonstrate the potential interest of including sequential dependence
in the implementation of an ICAMM classifier. Thus, a more accurate detection of
arousals in the hypnogram will help to the medical diagnosis of sleep disorders.
1.3.3.5 Discovering of Learning Styles in E-Learning
This application introduces the use of ICA as a data mining technique to extract
patterns of academic learning styles from a virtual campus. The analysis is based
on a well-known academic model (Felder's model) that classifies the learning style
of students according to the ways in which they receive and process information
[ 90 ]. Mixca is configured to estimate one ICA. It is demonstrated that the mixing
matrix could be used to associate dimensions of learning and web learning
activities from a huge amount (more than 2.3 millions of records) of historical web
log data. The learning styles detected automatically were consistent with the
courses and teaching methodologies of the e-learning campus. The results obtained
by ICA integrated with the results obtained by clustering and decision trees allow
an academic action outline for improving the learning process to be designed.
1.4 Overview
The thesis is presented in eight chapters. Excluding the Introduction (this chapter)
and Conclusions and Future Directions ( Chap. 8 ) , the thesis can be divided into
three parts: background ( Chap. 2 ) , theoretical contributions ( Chaps. 3 and 4 ), and
new applications ( Chaps. 5 , 6 and 7 ). Chapter 2 deals with theoretical foundations
about ICA and ICA mixtures. The performance of the proposed methods is
compared with selected algorithms that are representative of the many different
types of ICA and ICAMM algorithms that exist in the literature [ 13 - 18 ]. In order
to assist with these comparisons, a review of the selected algorithms is included.
The central contribution that this thesis makes is in Chap. 3 , which formulates a
general framework for modelling mixtures of ICAs. This chapter introduces the
ICA mixture model where each observation vector corresponds to a class which is
defined by an ICA model. Then the problem is stated as how to estimate the
mixture matrix and the bias terms for each class. This problem is approached by a
log-likelihood cost function of the unknowns, and an optimization procedure using
natural gradient algorithm is formulated. Four new features are included in the
resulting method for learning the ICA mixtures: non-parametric estimation of the