Search

The principal observed progressive swim types of sperm cells are linear mean and circular swim. Using motility characteristic parameters produced by CASA systems, we perform a parameter subset search to produce distinct clusters of the different swim types. For this task, the artificial bee colony algorithm (an iterative search algorithm modeled after the collective behavior of bees) and the well-studied k-means clustering algorithm were used on simulated and human sperm swim data. The result is distinct clusters with features of each types of swim. The clustering approach displays potential as a tool for automated sperm swim subpopulation analysis.

A hybrid feature selection method is proposed for classification in small sample size data sets.The filter step is based on instance learning taking advantage of the small sample size of data.A few candidate feature subsets are generated since their number corresponds to the number of instances.Cooperative feature subset search is proposed with a classifier algorithm for the wrapper step.The proposed method improves classification accuracy and stability of feature selection. The problem of selecting the most useful features from thousands of candidates in a low sample size data set arises in many areas of modern sciences. Feature subset selection is a key problem in such data mining classification tasks. In practice, it is very common to use filter methods. However, they ignore the correlations between genes which are prevalent in gene expression data. On the other hand, standard wrapper algorithms cannot be applied because of their complexity. Additionally, existing methods are not specially conceived to handle the small sample size of the data which is one of the main causes of feature selection instability. In order to deal with these issues, we propose a new hybrid, filter wrapper, approach based on instance learning. Its main challenge is that it converts the problem of the small sample size to a tool that allows choosing only a few subsets of features in a filter step. A cooperative subset search, CSS, is then proposed with a classifier algorithm to represent an evaluation system of wrappers. Our method is experimentally tested and compared with state-of-the-art algorithms based on several high-dimensional low sample size cancer datasets. Results show that our proposed approach outperforms other methods in terms of accuracy and stability of the selected subset.

For classifier ensemble systems, diversity is considered as an important factor for good generalization. In this paper, a classifier ensemble algorithm with relevance-based feature subset selection for classification is proposed. Firstly, a combined maximal class relevance and minimal feature relevance criterion is presented to evaluate candidate features, and to search diverse feature subspaces for classifier ensemble. And then bootstrap sampling is implemented on the feature subspaces for diverse training sets. Finally, the classifiers are trained on diverse data subspaces selected by feature subset search method with subsequent bootstrap. The experimental results show that the algorithm achieves high classification accuracies with small feature subspaces, which lead to a compact and effective ensemble system.

Previous studies have demonstrated that observers can search through a subset of items carrying a minority feature to find a conjunction target (Sobel & Cave, 2002). We examined whether subset search takes place when participants have less specific foreknowledge of the target (when the target is one of two possible items), measuring eye movements as well as reaction times. When there were unequal ratios of distractors, fixations were initially directed to the small subset. These initial eye movements were often directed between items with the same feature, suggesting guidance from pooled feature values. There was stronger guidance within color- than orientation-defined groups, although the features were balanced for salience. The results suggest that grouping of items by color operates more globally than grouping in orientation.

A novel method for measuring protein pocket similarity was devised, using only the alpha carbon positions of the pocket residues. Pockets were compared pairwise using an exhaustive three-dimensional Calpha common subset search, grouping residues by physicochemical properties. At least five Calpha matches were required for each hit, and distances between corresponding points were fit to an Extreme Value Distribution resulting in a probabilistic score or likelihood for any given superposition. A set of 85 structures from 13 diverse protein families was clustered based on binding sites alone, using this score. It was also successfully used to cluster 25 kinases into a number of subfamilies. Using a test kinase query to retrieve other kinase pockets, it was found that a specificity of 99.2% and sensitivity of 97.5% could be achieved using an appropriate cutoff score. The search itself took from 2 to 10 min on a single 3.4 GHz CPU to search the entire Protein Data Bank (133 800 pockets), depending on the number of hits returned.

In the present paper, we report two experiments out of series of studies designed to examine various aspects of visual search for icons of differing spatial frequencies. Specifically, the present experiments explore whether there exists a search asymmetry between high and low spatial frequency icons (A amongst B > B amongst A), and whether observers can limit their search to the relevant set of items in a display containing both types of icons. Our results show that a classic search asymmetry does not exist for spatial frequency; that, rather, both types of targets ‘pop out’; that search for a high spatial frequency target amongst high spatial frequency distractors is less efficient than search for a low spatial frequency target amongst low spatial frequency distractors; and that observers are partially able to limit their search to the relevant subset in mixed displays. Implications for the design of touch screen user interfaces are discussed.

Biopsy remains the gold standard for the diagnosis of chronic liver diseases. However, the variability in the diagnostic between readers leads to define a method to objectively describe histologic tissue. A complete framework has been implemented to analyze images of any tissue. Based on subset selection and feature ranking approaches, a feature selection computes the most relevant subset of descriptors in terms of classification from a wide initial list of descriptors. In comparison with equivalent methods, this implementation can find lists of descriptors which are significantly shorter for an equivalent accuracy. Furthermore, it enables the classification of slides using combinations of global and local measurements. The results have pointed that it could reach an accuracy of 90.5% (ROC-AUC=81.1%) in a human liver fibrosis grading approach by selecting 3 of the 457 global and local descriptors. The feature ranking approach gave less accurate subsets than the subset selection.

We consider a rough-set-inspired framework for deriving feature subset ensembles from data. Each of feature subsets yields a single classifier, basically by generating its corresponding if-then decision rules from the training data. Feature subsets are extracted according to a simple randomized algorithm, following the filter (rather than wrapper or embedded) methodology. Classifier ensemble is built from single classifiers by defining aggregation laws on top of decision rules. We investigate whether rough-set-inspired methods can help in the steps of formulating feature subset optimization criteria, feature subset search heuristics, and the strategies of voting among classifiers. Comparing to our previous research, we pay a special attention to synchronization of the filter-based criteria for feature subset selection and extraction of rules basing on the obtained feature subsets. The overall framework is not supposed to produce the best-ever classification results, unless it is extended by some additional techniques known from the literature. Our major goal is to illustrate in a possibly simplistic way some general interactions between the above-mentioned criteria.

We studied the involvement of the right parietal cortex in visual conjunction search, where two features are present in the array and spatial attention and feature binding is required, and in subset search, where two features are also present but only one of them is needed in order to group stimuli together (the subset) and allow parallel processing without the need for feature binding. Six patients with right parietal lobe lesions, six age-matched controls, and three control patients with left parietal lesions were tested on these two tasks. Patients with right parietal lesions were significantly slower than normal controls in the conjunction task, especially for target-absent trials. In the subset condition, neither normal control subjects nor patients with left parietal damage showed a difference between target-present and target-absent trials whereas right parietal patients showed a significant difference between target-present and target-absent responses. The results suggest a role for the right parietal cortex in shifting attention to the next stimulus once binding of features has taken place or selecting spatial areas containing the desired feature in a subset search, but that parietal cortex is not required for binding the features of the object.

The combinatorial search problem arising in feature selection in high dimensional spaces is considered. Recently developed techniques based on the classical sequential methods and the (l, r) search called Floating search algorithms are compared against the Genetic approach to feature subset search. Both approaches have been designed with the view to give a good compromise between efficiency and effectiveness for large problems. The purpose of this paper is to investigate the applicability of these techniques to high dimensional problems of feature selection. The aim is to establish whether the properties inferred for these techniques from medium scale experiments involving up to a few tens of dimensions extend to dimensionalities of one order of magnitude higher. Further, relative merits of these techniques vis-a-vis such high dimensional problems are explored and the possibility of exploiting the best aspects of these methods to create a composite feature selection procedure with superior properties is considered.