This paper investigates application of novel Bidirectional Data Partitioning Technique (BDP) to cancer survival analysis. Author has developed this technique for classification problems with unstable feature relevance and SEER Cancer Data illustrates this machine learning concept. BDP is applied for survival analysis in order to find groups of patients with different key factors that determine survival time. BDP operates by weights assigned to instance-feature tuples. A measure of class separability is used as a criterion in finding weights. Weights are then used within clustering in order to find clusters (subgroups of patients) with associated feature importance profiles (scored survival factors). Component models of an ensemble are built after agglomerative merging of subgroups. Because different clustering techniques typically yield different results, accuracy of an ensemble is used to establish final groups. Factors of survival time that are crucial in different situations define survival groups according to a dissimilarity principle. The results establish a base for additional epidemiological surveys and studies.

In this paper, analysis of class separability measures is performed in attempt to relate their descriptive abilities to geometrical properties of classification problems in presence of irrelevant features. The study is performed on synthetic and benchmark data with known irrelevant features and other characteristics of interest, such as class boundaries, shapes, margins between classes, and density. The results have shown that some measures are individually informative, while others are less reliable and only can provide complimentary information. Classification problem complexity measurements on selected data sets are made to gain additional insights on the obtained results.

The approach to construct predictive models of heterogeneous data is based on decomposition of a classification problem into subproblems. Applicability of this approach depends on the success in discovering the homogeneous regions in data and their coverage by the local predictive models. The importance Profile Angle (iPA) may provide an additional indication of heterogeneity considering profiles of feature importance in subproblems. in this paper iPA is evaluated on several variations of heterogeneity. The experimental study on the data sets with known data characteristics related to heterogeneity has shown that iP A is applicable when the feature merit measures are identified adequately. indication of heterogeneity provided by iP A has been verified via the gains in classification accuracy obtained in subproblems after decomposition.

We present a complete system architecture for fully automated markerless augmented reality (AR). The system constructs a sparse metric model of the real-world environment, provides interactive means for specifying the pose of a virtual object, and performs model-based camera tracking with visually pleasing augmentation results. Our approach does not require camera pre-calibration, prior knowledge of scene geometry, manual initialization of the tracker or placement of special markers. Robust tracking in the presence of occlusions and scene changes is achieved by using highly distinctive natural features to establish image correspondences.

Diagnostics of hearing impairments is a non-trivial problem for data mining techniques. The state of hearing can be described via a measurement of polymorphic disorders in the voice structure that are secondary to restricted auditory control. The diagnostic voice analysis determines voice descriptors that can be used for marginal estimation of the state of hearing. This problem is hard for most of the predictive data mining methods. The presence of strongly correlated and redundant information in the set of voice descriptors might be one reason for the low prediction accuracy. In this paper, different feature selection techniques are evaluated by their ability to raise the prediction accuracy by discarding irrelevant and redundant voice descriptors when modeling the dependency between functional changes within a phonatory organ and restricted auditory control. As the result of the prediction varies for different prediction methods, the applicability of certain feature selection technique is considered with respect to the prediction method and evaluated as a feature selection strategy.

Real-world medical data is often heterogeneous, containing many cases and features, each of which requires different a type of processing. Generally, this means that the subsets of relevant features are different for various cases. The set of voice descriptors in the problem of hearing impairment diagnosis is an example of such a heterogeneous domain. Ensemble feature selection techniques are adopted to take into account the data heterogeneity. This paper analyses the applicability of various feature selection approaches in diagnosing hearing impairments in the context of an ensemble classification. Ensemble feature selection produces multiple classifiers for this domain, based on feature subsets derived by different feature selection approaches. In particular, we are interested in performing feature selection for each particular case, taking into consideration any hidden heterogeneity in the data. We use real-world clinical hearing impairment data and compare ensemble classification to the single-classifier technique.

Current electronic data repositories contain enormous amounts of data, especially in medical domains, where data is often feature-space heterogeneous, so that different features have different importance in different sub-areas of the whole space. In this paper, we suggest a technique that searches for a strategic splitting of the feature space, identifying the best subsets of features for each instance. Our technique is based on the wrapper approach, where a classification algorithm is used as the evaluation function to differentiate between several feature subsets. We apply a recently developed technique for the dynamic integration of classifiers and use decision trees. For each test instance, we consider only those feature combinations that include features that are present in the path taken by the test instance in the decision tree. We evaluate our technique on medical data sets from the UCI machine learning repository. The experiments show that local feature selection is often advantageous in comparison with feature selection on the whole space.