Conclusions - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 176

dimensions of the learning styles of the students. Thus, the observations were
linear combinations of those styles through the use of the facilities at the
virtual campus. Significant learning styles were detected for students of
courses with grades. Lack of assessment in the courses did not allow learning
styles to be identified. This confirmed a known pedagogical principle, that
expected evaluation biases how the student learns.
8.3 Future Work
There are several open research topics that will improve the proposed methods.
These topics are listed below:
8.3.1 Improvement of ICAMM Extensions
• Research on strategies to improve the parameter initialization. This is an
important issue since the starting point of the process of learning determines
the convergence of the algorithm. Regularization and penalization techniques
can be applied to avoid convergence to local minima.
• Research on the residual dependence after convergence in classification and
hierarchical clustering. The independence of the hidden variables is difficult
to find in nature. Thus, information of class membership or posterior prob-
abilities (i.e., the probability of every class conditioned to the feature of the
observation vector) can be used to model fuzzy rules that reflect the
dependency in the data mixture. In addition, dependency measures could be
used to develop novelty detection procedures.
• Incorporation of the use of priors in the estimation of the source densities.
This can be useful for particular applications in which features of the
objective signals are known (statistical or spectral features such as bandwidth
range, and kurtosis type). This is one of the current lines of research in ICA.
However, the advantage of using priors to lead the algorithm to the objective
is also a restriction that diminishes the blindness principle.
• Development of techniques to detect and process outliers. In some applica-
tions, the outliers are not strange data to be removed, but they are the
interesting novelty to be found. The hierarchical classification of the data
offers several possibilities to detect outliers such as the analysis of particular
patterns in the manner in which the binary tree of clustering is built.