Search

In this paper, we introduce a new hybrid filter-wrapper method for supervised feature selection, based on the Laplacian Score ranking combined with a wrapper strategy. We propose to rank features with the Laplacian Score to reduce the search space, and then we use this order to find the best feature subset. We compare our method against other based on ranking feature selection methods, namely, Information Gain Attribute Ranking, Relief, Correlation-based Feature Selection, and additionally we include in our comparison a Wrapper Subset Evaluation method. Empirical results over ten real-world datasets from the UCI repository show that our hybrid method is competitive and outperforms in most of the cases to the other feature selection methods used in our experiments. [ABSTRACT FROM AUTHOR]

Several algorithms for induction of decision trees have been developed to solve problems with large datasets, however some of them have spatial and/or runtime problems using the whole training sample for building the tree and others do not take into account the whole training set. In this paper, we introduce a new algorithm for inducing decision trees for large numerical datasets, called IIMDT, which builds the tree in an incremental way and therefore it is not necesary to keep in main memory the whole training set. A comparison between IIMDT and ICE, an algorithm for inducing decision trees for large datasets, is shown. [ABSTRACT FROM AUTHOR]

The k nearest neighbor (k-NN) classifier has been a widely used nonparametric technique in Pattern Recognition. In order to decide the class of a new prototype, the k-NN classifier performs an exhaustive comparison between the prototype to classify (query) and the prototypes in the training set T. However, when T is large, the exhaustive comparison is expensive. To avoid this problem, many fast k-NN algorithms have been developed. Some of these algorithms are based on Approximating-Eliminating search. In this case, the Approximating and Eliminating steps rely on the triangle inequality. However, in soft sciences, the prototypes are usually described by qualitative and quantitative features (mixed data), and sometimes the comparison function does not satisfy the triangle inequality. Therefore, in this work, a fast k most similar neighbour classifier for mixed data (AEMD) is presented. This classifier consists of two phases. In the first phase, a binary similarity matrix among the prototypes in T is stored. In the second phase, new Approximating and Eliminating steps, which are not based on the triangle inequality, are presented. The proposed classifier is compared against other fast k-NN algorithms, which are adapted to work with mixed data. Some experiments with real datasets are presented. [ABSTRACT FROM AUTHOR]

In supervised classification, the object selection or instance selection is an important task, mainly for instance-based classifiers since through this process the time in training and classification stages could be reduced. In this work, we propose a new mixed data object selection method based on clustering and border objects. We carried out an experimental comparison between our method and other object selection methods using some mixed data classifiers. [ABSTRACT FROM AUTHOR]

In this work, a fast k most similar neighbor (k-MSN) classifier for mixed data is presented. The k nearest neighbor (k-NN) classifier has been a widely used nonparametric technique in Pattern Recognition. Many fast k-NN classifiers have been developed to be applied on numerical object descriptions, most of them based on metric properties to avoid object comparisons. However, in some sciences as Medicine, Geology, Sociology, etc., objects are usually described by numerical and non numerical features (mixed data). In this case, we can not assume the comparison function satisfies metric properties. Therefore, our classifier is based on search algorithms suitable for mixed data and non-metric comparison functions. Some experiments and a comparison against other two fast k-NN methods, using standard databases, are presented. [ABSTRACT FROM AUTHOR]

Object selection is an important task for instance-based classifiers since through this process the time in training and classification stages could be reduced. In this work, we propose a new method based on clustering which tries to find border objects that contribute with useful information allowing to the classifier discriminating between classes. An experimental comparison of our method, the CLU method based on clustering, and the DROP methods, is presented. [ABSTRACT FROM AUTHOR]

The object selection is an important task for instance-based classifiers since through this process the size of a training set could be reduced and then the runtimes in both classification and training steps would be reduced. Several methods for object selection have been proposed but some methods discard relevant objects for the classification step. In this paper, we propose an object selection method which is based on the idea of sequential floating search. This method reconsiders the inclusion of relevant objects previously discarded. Some experimental results obtained by our method are shown and compared against some other object selection methods. [ABSTRACT FROM AUTHOR]

The nearest neighbor (NN) classifier has been a widely used technique in pattern recognition because of its simplicity and good behavior. To decide the class of a new object, the NN classifier performs an exhaustive comparison between the object to classify and the training set T. However, when T is large, the exhaustive comparison is very expensive and sometimes becomes inapplicable. To avoid this problem, many fast NN algorithms have been developed for numerical object descriptions, most of them based on metric properties to avoid comparisons. However, in some sciences as Medicine, Geology, Sociology, etc., objects are usually described by numerical and non numerical attributes (mixed data). In this case, we can not assume the comparison function satisfies metric properties. Therefore, in this paper a fast most similar object classifier based on search methods suitable for mixed data is presented. Some experiments using standard databases and a comparison with other two fast NN methods are presented. [ABSTRACT FROM AUTHOR]

Typical testors are a useful tool to do feature selection in supervised classification problems with mixed incomplete data. However, the complexity of computing all typical testors of a training matrix has an exponential growth with respect to the number of columns in the matrix. For this reason different approaches like heuristic algorithms, parallel and distributed processing, have been developed. In this paper, we present a configurable custom architecture for the efficient identification of testors from a given input matrix. The architectural design is based on a brute force approach that is suitable for high populated input matrixes. The architecture has been designed to deal with parallel processing and can be configured for any size of matrix. The architecture is able to evaluate if a vector is a testor of the matrix in a single clock cycle. The architecture has been implemented on a Field Programmable Gate Array (FPGA) device. Results show that it provides runtime improvements over software implementations running on state-of-the-art processors. FPGA implementation results are presented and implications to the field of pattern recognition discussed. [ABSTRACT FROM AUTHOR]

The k-means algorithm is the most studied and used tool for solving the clustering problem when the number of clusters is known a priori. Nowadays, there is only one conceptual version of this algorithm, the conceptual k-means algorithm. One characteristic of this algorithm is the use of generalization lattices, which define relationships among the feature values. However, for many applications, it is difficult to determine the best generalization lattices; moreover there are not automatic methods to build the lattices, thus this task must be done by the specialist of the area in which we want to solve the problem. In addition, this algorithm does not work with missing data. For these reasons, in this paper, a new conceptual k-means algorithm that does not use generalization lattices to build the concepts and allows working with missing data is proposed. We use complex features for generating the concepts. The complex features are subsets of features with associated values that characterize objects of a cluster and at the same time not characterize objects from other clusters. Some experimental results obtained by our algorithm are shown and they are compared against the results obtained by the conceptual k-means algorithm. [ABSTRACT FROM AUTHOR]

Edition is an important and useful task in supervised classification specifically for instance-based classifiers because edition discards from the training set those useless or harmful objects for the classification accuracy and it helps to reduce the size of the original training sample and to increase both the classification speed and accuracy. In this paper, we propose two edition schemes that combine edition methods and sequential search for instance selection. In addition, we present an empirical comparison between these schemes and some other edition methods. [ABSTRACT FROM AUTHOR]

The edition process is an important task in supervised classification because it helps to reduce the size of the training sample. On the other hand, Instance-Based classifiers store all the training set indiscriminately, which in almost all times, contains useless or harmful objects, for the classification process. Therefore it is important to delete unnecessary objects to increase both classification speed and accuracy. In this paper, we propose an edition method based on sequential search and we present an empirical comparison between it and some other decremental edition methods. [ABSTRACT FROM AUTHOR]