@PhDThesis{Krizek-CAK-2008-31,
  UPDATE  = { 2008-11-27 },
  author =       {K{\v r}{\'\i}{\v z}ek, Pavel},
  supervisor =   {Hlav\'a\v{c}, V{\'a}clav and Kittler, Josef},
  title =        {Feature selection: Stability, algorithms, and evaluation},
  school =       {Center for Machine Perception, K13133 FEE Czech Technical
                  University},
  address =      {Prague, Czech Republic},
  year =         {2008},
  month =        {November},
  day =          {14},
  type =         {{PhD Thesis K333--31/08, CTU--CMP--2008--16}},
  pages =        {120},
  figures =      {37},
  authorship =   {100},
  psurl =        {[Krizek-TR-2008-16.pdf]},
  project =      {1M0567, ICT-215078 DIPLECS},
  annote = {In the framework of statistical pattern recognition,
    feature selection is an important step aiming to extract the most
    important discriminatory information for classification and
    compile it concisely into a pattern vector of a lower
    dimensionality. The motivation for applying feature selection is
    multifacetted. Features can be expensive to acquire. The costs
    include measurement acquisition, data preprocessing, data transfer
    and storage, and computational costs. Furthermore,
    high-dimensional problems need more samples for training to
    achieve a good generalization capability of a classifier (ie, the
    curse of dimensionality). Reduced dimensionality of the pattern
    vector can also help, in applications like medicine or genetic
    engineering, to gain better understanding of a given problem. The
    main goal of this thesis is to design a classifier independent
    (filter-based) feature selection method that would allow the merit
    of individual features to be assessed from one-dimensional
    projections of the data. However, we have encountered several
    other problems during our research which had to be addressed
    first. We therefore also discuss the topics concerning data
    re-sampling and error rate estimation, the evaluation ethodology
    applicable to feature selection, and the highly neglected topic of
    feature set stability in the context of feature selection. We
    propose three filter-based feature selection algorithms. The first
    two methods can be considered as an unsupervised pre-filtering
    step for more advanced feature selection algorithms. The third
    method exploits the information conveyed by the evolution of
    weights assigned to the the training samples similarly to the
    Adaboost algorithm. In experiments with synthetic and real-world
    data we show that features selected by the proposed methods
    improve the performance of commonly employed learning
    methods. Moreover, the selected features may even outperform the
    features selected by the current state-of-the-art techniques. One
    of the main contributions of this thesis is a theoretically
    justified measure for assessing the stability of feature selection
    results in the presence of random perturbations in the input
    training data. Among other results, our findings suggest, that the
    best stability and performance of the features selected with
    classifier dependent feature selection methods is achieved if the
    criterion estimation is accomplished using the repeated two-fold
    cross-validation. },
  keywords = {Statistical pattern recognition, feature selection},
}