Intelligent Data Analysis: An International Journal

We propose a statistical approach based on a supervised framework for reducing the dimensionality of the feature space when characterizing and classifying spatial Regions of Interest (ROIs). Our approach employs the statistical techniques of Bootstrapping simulation, Bayesian Inference and Markov Chain Monte Carlo (MCMC), to select the most informative features according to their discriminative power across distinct classes of data. This reduces the dimensionality of the initial feature space and also improves the classification of the ROIs, since features providing irrelevant information with respect to class membership are discarded. We also introduce a weighted Euclidean Distance designed to effectively classify the ROIs. We evaluate the proposed technique using experiments that involve synthetic spatial regions and real ROIs extracted from medical images. We demonstrate its effectiveness in classification experiments (using established classifiers) and in similarity searches. We also test its scalability on large datasets. Our approach is comparable with or better than other major competitors. We achieve an accuracy of 87% on classifying ROIs in brain images. These results are an improvement of previously reported classification experiments, and show the effect of reducing the dimensionality of the initial feature space.

We present a method to extract a time series (Number of Active Requests (NAR)) from web cache logs which serves as a transport level measurement of internet traffic. This series also reflects the performance or Quality of Service of a web cache. It has long-memory properties but is not self-similar and does not have a heavy-tailed distribution.

However, the long-memory and autocorrelation structure of NAR are preserved through aggregation, that is the aggregated series has similar statistical properties to the original one. We call this property aggregation similarity.

Aggregation similarity is a very useful property, which makes management of large data sets easier and speeds up the asymptotic properties of time series.

The estimation of mixture models has been proposed for quite some time as an approach for cluster analysis. Several variants of the Expectation-Maximization algorithm are currently available for this purpose. Estimation of mixture models simultaneously allows the determination of the number of clusters and yields distributional parameters for clustering base variables. There are several information criteria that help to support the selection of a particular model or clustering structure. However, a question remains concerning the selection of specific criteria that may be more suitable for particular applications. In the present work we analyze the relationship between the performance of information criteria and the type of measurement of clustering variables. In order to study this relationship we perform the analysis of forty-two data sets with known clustering structure and with clustering variables that are categorical, continuous and mixed type. We then compare eleven information-based criteria in their ability to recover the data sets' clustering structures. As a result, we select AIC3, BIC and ICL-BIC criteria as the best candidates for model selection that refers to models with categorical, continuous and mixed type clustering variables, respectively.

DNA micro-arrays provide thousands of genomic expressions on the same subject. A main issue is then to find the subset of genes whose degeneration is responsible of a certain type of cancer. In this paper, starting from a paradigmatic classification problem of two kinds of Leukaemia, we discuss the use of data-mining techniques in such a context. Particular attention is devoted not only to the classification method but also to all the data analysis steps including data pre-processing and information retrieval.

Modeling with real-world data is often plagued with the problem of missing values, limiting the applicability and validity of the developed model. Several algorithms exist in the literature to facilitate the analysis of incomplete data by imputing missing values. However, their imputation accuracy and practical applicability have not been systematically compared and studied. This makes the choice of appropriate imputation method difficult. The focus of this paper is to conduct an exploratory analysis of the popular missing data imputation algorithms. A new imputation algorithm based on clustering is also developed and demonstrated to be useful in a variety of ways to improve the efficiency of imputing missing values. These algorithms are benchmarked using datasets with significantly varying statistical properties. Based on the empirical results and theoretical analysis, a set of guidelines are proposed to assist in the selection of an appropriate imputation algorithm for a specific application. Finally these guidelines are used in a process modeling case study that involves the analysis of the design of an atomizer. It was observed that the imputed values are qualitatively valid thus providing evidence for the appropriateness of the proposed guidelines.

The computation of compact and meaningful representations of high dimensional sensor data has recently been addressed through the development of Nonlinear Dimensional Reduction (NLDR) algorithms. The numerical implementation of spectral NLDR techniques typically leads to a symmetric eigenvalue problem that is solved by traditional batch eigensolution algorithms. The application of such algorithms in real-time systems necessitates the development of sequential algorithms that perform feature extraction online. This paper presents an efficient online NLDR scheme, Sequential-Isomap, based on incremental singular value decomposition (SVD) and the Isomap method. Example simulations demonstrate the validity and significant potential of this technique in real-time applications such as autonomous systems.