Advances in Data Analysis and Classification最新文献

One of the most frequently used algorithms for clustering data with both numeric and categorical variables is the k-prototypes algorithm, an extension of the well-known k-means clustering. Gower’s distance denotes another popular approach for dealing with mixed-type data and is suitable not only for numeric and categorical but also for ordinal variables. In the paper a modification of the k-prototypes algorithm to Gower’s distance is proposed that ensures convergence. This provides a tool that allows to take into account ordinal information for clustering and can also be used for large data. A simulation study demonstrates convergence, good clustering results as well as small runtimes.

Canonical variate analysis (CVA) entails a two-sided eigenvalue decomposition. When the number of groups, J, is less than the number of variables, p, at most (J-1) eigenvalues are not exactly zero. A CVA biplot is the simultaneous display of the two entities: group means as points and variables as calibrated biplot axes. It follows that with two groups the group means can be exactly represented in a one-dimensional biplot but the individual samples are approximated. We define a criterion to measure the quality of representing the individual samples in a CVA biplot. Then, for the two-group case we propose an additional dimension for constructing an optimal two-dimensional CVA biplot. The proposed novel CVA biplot maintains the exact display of group means and biplot axes, but the individual sample points satisfy the optimality criterion in a unique simultaneous display of group means, calibrated biplot axes for the variables, and within group samples. Although our primary aim is to address two-group CVA, our proposal extends immediately to an optimal three-dimensional biplot when encountering the equally important case of comparing three groups in practice.

Among different functional data analyses, clustering analysis aims to determine underlying groups of curves in the dataset when there is no information on the group membership of each curve. In this work, we develop a novel variational Bayes (VB) algorithm for clustering and smoothing functional data simultaneously via a B-spline regression mixture model with random intercepts. We employ the deviance information criterion to select the best number of clusters. The proposed VB algorithm is evaluated and compared with other methods (k-means, functional k-means and two other model-based methods) via a simulation study under various scenarios. We apply our proposed methodology to two publicly available datasets. We demonstrate that the proposed VB algorithm achieves satisfactory clustering performance in both simulation and real data analyses.

The paper discusses eleven internal evaluation criteria that can be used in the area of hierarchical clustering of categorical data. The criteria are divided into two distinct groups based on how they treat the cluster quality: variability- and distance-based. The paper follows three main aims. The first one is to compare the examined criteria regarding their mutual similarity and dependence on the clustered datasets’ properties and the used similarity measures. The second one is to analyze the relationships between internal and external cluster evaluation to determine how well the internal criteria can recognize the original number of clusters in datasets and to what extent they provide comparable results to the external criteria. The third aim is to propose two new variability-based internal evaluation criteria. In the experiment, 81 types of generated datasets with controlled properties are used. The results show which internal criteria can be recommended for specific tasks, such as judging the cluster quality or the optimal number of clusters determination.

We present a set of algorithms implementing multidimensional scaling (MDS) for large data sets. MDS is a family of dimensionality reduction techniques using a (n times n) distance matrix as input, where n is the number of individuals, and producing a low dimensional configuration: a (ntimes r) matrix with (r<<n). When n is large, MDS is unaffordable with classical MDS algorithms because their extremely large memory and time requirements. We compare six non-standard algorithms intended to overcome these difficulties. They are based on the central idea of partitioning the data set into small pieces, where classical MDS methods can work. Two of these algorithms are original proposals. In order to check the performance of the algorithms as well as to compare them, we have done a simulation study. Additionally, we have used the algorithms to obtain an MDS configuration for EMNIST: a real large data set with more than 800000 points. We conclude that all the algorithms are appropriate to use for obtaining an MDS configuration, but we recommend to use one of our proposals, since it is a fast algorithm with satisfactory statistical properties when working with big data. An R package implementing the algorithms has been created.

Multi-view stacking is a framework for combining information from different views (i.e. different feature sets) describing the same set of objects. In this framework, a base-learner algorithm is trained on each view separately, and their predictions are then combined by a meta-learner algorithm. In a previous study, stacked penalized logistic regression, a special case of multi-view stacking, has been shown to be useful in identifying which views are most important for prediction. In this article we expand this research by considering seven different algorithms to use as the meta-learner, and evaluating their view selection and classification performance in simulations and two applications on real gene-expression data sets. Our results suggest that if both view selection and classification accuracy are important to the research at hand, then the nonnegative lasso, nonnegative adaptive lasso and nonnegative elastic net are suitable meta-learners. Exactly which among these three is to be preferred depends on the research context. The remaining four meta-learners, namely nonnegative ridge regression, nonnegative forward selection, stability selection and the interpolating predictor, show little advantages in order to be preferred over the other three.

Clustering ensemble combines several fundamental clusterings with a consensus function to produce the final clustering without gaining access to data features. The quality and diversity of a vast library of base clusterings influence the performance of the consensus function. When a huge library of various clusterings is not available, this function produces results of lower quality than those of the basic clustering. The expansion of diverse clusters in the collection to increase the performance of consensus, especially in cases where there is no access to specific data features or assumptions in the data distribution, has still remained an open problem. The approach proposed in this paper, Clustering Ensemble Extraction, considers the similarity criterion at the cluster level and places the most similar clusters in the same group. Then, it extracts new clusters with the help of the Extracting Clusters Algorithm. Finally, two new consensus functions, namely Cluster-based extracted partitioning algorithm and Meta-cluster extracted algorithm, are defined and then applied to new clusters in order to create a high-quality clustering. The results of the empirical experiments conducted in this study showed that the new consensus function obtained by our proposed method outperformed the methods previously proposed in the literature regarding the clustering quality and efficiency.

Abstract

Finite mixtures of regressions (FMRs) are powerful clustering devices used in many regression-type analyses. Unfortunately, real data often present atypical observations that make the commonly adopted normality assumption of the mixture components inadequate. Thus, to robustify the FMR approach in a matrix-variate framework, we introduce ten FMRs based on the matrix-variate t and contaminated normal distributions. Furthermore, once one of our models is estimated and the observations are assigned to the groups, different procedures can be used for the detection of the atypical points in the data. An ECM algorithm is outlined for maximum likelihood parameter estimation. By using simulated data, we show the negative consequences (in terms of parameter estimates and inferred classification) of the wrong normality assumption in the presence of heavy-tailed clusters or noisy matrices. Such issues are properly addressed by our models instead. Additionally, over the same data, the atypical points detection procedures are also investigated. A real-data analysis concerning the relationship between greenhouse gas emissions and their determinants is conducted, and the behavior of our models in the presence of heterogeneity and atypical observations is discussed.