Journal of Classification最新文献

Compositional data have peculiar characteristics that pose significant challenges to traditional statistical methods and models. Within this framework, we use a convenient mode parametrized Dirichlet distribution across multiple fields of statistics. In particular, we propose finite mixtures of unimodal Dirichlet (UD) distributions for model-based clustering and classification. Then, we introduce the contaminated UD (CUD) distribution, a heavy-tailed generalization of the UD distribution that allows for a more flexible tail behavior in the presence of atypical observations. Thirdly, we propose finite mixtures of CUD distributions to jointly account for the presence of clusters and atypical points in the data. Parameter estimation is carried out by directly maximizing the maximum likelihood or by using an expectation-maximization (EM) algorithm. Two analyses are conducted on simulated data to illustrate the effects of atypical observations on parameter estimation and data classification, and how our proposals address both aspects. Furthermore, two real datasets are investigated and the results obtained via our models are discussed.

In this paper, we propose TAWE (title assignment with word embedding), a new method to automatically assign titles to topics inferred from sets of documents. This method combines the results obtained from the topic modeling performed with, e.g., latent Dirichlet allocation (LDA) or other suitable methods and the word embedding representation of words in a vector space. This representation preserves the meaning of the words while allowing to find the most suitable word that represents the topic. The procedure is twofold: first, a cleaned text is used to build the LDA model to infer a desirable number of latent topics; second, a reasonable number of words and their weights are extracted from each topic and represented in n-dimensional space using word embedding. Based on the selected weighted words, a centroid is computed, and the closest word is chosen as the title of the topic. To test the method, we used a collection of tweets about climate change downloaded from some of the main newspapers accounts on Twitter. Results showed that TAWE is a suitable method for automatically assigning a topic title.

There is no, nor will there ever be, single best clustering algorithm. Nevertheless, we would still like to be able to distinguish between methods that work well on certain task types and those that systematically underperform. Clustering algorithms are traditionally evaluated using either internal or external validity measures. Internal measures quantify different aspects of the obtained partitions, e.g., the average degree of cluster compactness or point separability. However, their validity is questionable because the clusterings they endorse can sometimes be meaningless. External measures, on the other hand, compare the algorithms’ outputs to fixed ground truth groupings provided by experts. In this paper, we argue that the commonly used classical partition similarity scores, such as the normalised mutual information, Fowlkes–Mallows, or adjusted Rand index, miss some desirable properties. In particular, they do not identify worst-case scenarios correctly, nor are they easily interpretable. As a consequence, the evaluation of clustering algorithms on diverse benchmark datasets can be difficult. To remedy these issues, we propose and analyse a new measure: a version of the optimal set-matching accuracy, which is normalised, monotonic with respect to some similarity relation, scale-invariant, and corrected for the imbalancedness of cluster sizes (but neither symmetric nor adjusted for chance).

Many evaluations of binary classifiers begin by adopting a pair of indicators, most often sensitivity and specificity or precision and recall. Despite this, we lack a general, pan-disciplinary basis for choosing one pair over the other, or over one of four other sibling pairs. Related obscurity afflicts the choice between the receiver operating characteristic and the precision-recall curve. Here, I return to first principles to separate concerns and distinguish more than 50 foundational concepts. This allows me to establish six rules that allow one to identify which pair is correct. The choice depends on the context in which the classifier is to operate, the intended use of the classifications, their intended user(s), and the measurability of the underlying classes, but not skew. The rules can be applied by those who develop, operate, or regulate them to classifiers composed of technology, people, or combinations of the two.

Clustering longitudinal data for growth curve modelling is considered in this paper, where we aim to optimally estimate the underpinning unknown group partition matrix. Instead of following the conventional soft clustering approach, which assumes the columns of the partition matrix to have i.i.d. multinomial or categorical prior distributions and uses a regression model with the response following a finite mixture distribution to estimate the posterior distribution of the partition matrix, we propose an iterative partition and regression procedure to find the best partition matrix and the associated best growth curve regression model for each identified cluster. We show that the best partition matrix is the one minimizing a recently developed empirical Bayes information criterion (eBIC), which, due to the involved combinatorial explosion, is difficult to compute via enumerating all candidate partition matrices. Thus, we develop a Gibbs sampling method to generate a Markov chain of candidate partition matrices that has its equilibrium probability distribution equal the one induced from eBIC. We further show that the best partition matrix, given a priori the number of latent clusters, can be consistently estimated and is computationally scalable based on this Markov chain. The number of latent clusters is also best estimated by minimizing eBIC. The proposed iterative clustering and regression method is assessed by a comprehensive simulation study before being applied to two real-world growth curve modelling examples involving longitudinal data clustering.

Clustering, or unsupervised classification, is a task often plagued by outliers. Yet there is a paucity of work on handling outliers in clustering. Outlier identification algorithms tend to fall into three broad categories: outlier inclusion, outlier trimming, and post hoc outlier identification methods, with the former two often requiring pre-specification of the number of outliers. The fact that sample squared Mahalanobis distance is beta-distributed is used to derive an approximate distribution for the log-likelihoods of subset finite Gaussian mixture models. An algorithm is then proposed that removes the least plausible points according to the subset log-likelihoods, which are deemed outliers, until the subset log-likelihoods adhere to the reference distribution. This results in a trimming method, called OCLUST, that inherently estimates the number of outliers.

This paper introduces a novel supervised classification method based on dynamic clustering (DC) and K-nearest neighbor (KNN) learning algorithms, denoted DC-KNN. The aim is to improve the accuracy of a classifier by using a DC method to discover the hidden patterns of the apriori groups of the training set. It provides a partitioning of each group into a predetermined number of subgroups. A new objective function is designed for the DC variant, based on a trade-off between the compactness and separation of all subgroups in the original groups. Moreover, the proposed DC method uses adaptive distances which assign a set of weights to the variables of each cluster, which depend on both their intra-cluster and inter-cluster structure. DC-KNN performs the minimization of a suitable objective function. Next, the KNN algorithm takes into account objects by assigning them to the label of subgroups. Furthermore, the classification step is performed according to two KNN competing algorithms. The proposed strategies have been evaluated using both synthetic data and widely used real datasets from public repositories. The achieved results have confirmed the effectiveness and robustness of the strategy in improving classification accuracy in comparison to alternative approaches.

Data clustering is an important task in the field of data mining. In many real applications, clustering algorithms must consider the order of data, resulting in the problem of clustering sequential data. For instance, analyzing the moving pattern of an object and detecting community structure in a complex network are related to sequential data clustering. The constraint of the continuous region prevents previous clustering algorithms from being directly applied to the problem. A dynamic programming algorithm was proposed to address the issue, which returns the optimal sequential data clustering. However, it is not scalable and hence the practicality is limited. This paper revisits the solution and enhances it by introducing a greedy stopping condition. This condition halts the algorithm’s search process when it is likely that the optimal solution has been found. Experimental results on multiple datasets show that the algorithm is much faster than its original solution while the optimality gap is negligible.

The family of multiple scaled mixtures of multivariate normal (MSMN) distributions has been shown to be a powerful tool for modeling data that allow different marginal amounts of tail weight. An extension of the MSMN distribution is proposed through the incorporation of a vector of shape parameters, resulting in the skew multiple scaled mixtures of multivariate normal (SMSMN) distributions. The family of SMSMN distributions can express a variety of shapes by controlling different degrees of tailedness and versatile skewness in each dimension. Some characterizations and probabilistic properties of the SMSMN distributions are studied and an extension to finite mixtures thereof is also discussed. Based on a sort of selection mechanism, a feasible ECME algorithm is designed to compute the maximum likelihood estimates of model parameters. Numerical experiments on simulated data and three real data examples demonstrate the efficacy and usefulness of the proposed methodology.

For supervised classification we propose to use restricted multidimensional unfolding in a multinomial logistic framework. Where previous research proposed similar models based on squared distances, we propose to use usual (i.e., not squared) Euclidean distances. This change in functional form results in several interpretational advantages of the resulting biplot, a graphical representation of the classification model. First, the conditional probability of any class peaks at the location of the class in the Euclidean space. Second, the interpretation of the biplot is in terms of distances towards the class points, whereas in the squared distance model the interpretation is in terms of the distance towards the decision boundary. Third, the distance between two class points represents an upper bound for the estimated log-odds of choosing one of these classes over the other. For our multinomial restricted unfolding, we develop and test a Majorization Minimization algorithm that monotonically decreases the negative log-likelihood. With two empirical applications we point out the advantages of the distance model and show how to apply multinomial restricted unfolding in practice, including model selection.