Search

Abstract Multi-view unsupervised feature selection has been proven to be an effective approach to reduce the dimensionality of multi-view data. One of its key issues is how to exploit the underlying common structures across different views. In this paper, we propose a consensus learning guided multi-view unsupervised feature selection method, which embeds multi-view feature selection into a non-negative matrix factorization based clustering with sparse constrain. The proposed method learns latent feature matrices from all the views, and optimizes a consensus matrix such that the difference between the cluster indicator matrix of each view and the consensus matrix is minimized. The parameters for balancing the weights of different views are automatically adjusted, and a sparse constraint is imposed on the latent feature matrices to perform feature selection. After that, we design an effective iterative algorithm to solve the resultant optimization problem. Extensive experiments have been conducted on six publicly multi-view datasets, and the results demonstrate that the proposed algorithm outperforms several other state-of-the-art single view and multi-view unsupervised feature selection methods in terms of clustering tasks, validating the effectiveness of the proposed multi-view unsupervised feature selection method. The source code of our algorithm will be available on our on-line page: http://tangchang.net/. [ABSTRACT FROM AUTHOR]

The aim of the paper is to explore the evolution of food diets in 40 European countries according to the common European policies and guidelines on healthy diets. To this end, an innovative clustering method, called CoClust, has been adopted. By means of the copula function, this algorithm is able to find clusters based on the complex multivariate dependence structure of the data generating process, overcoming the limits of classical approaches that cope with only linear bivariate relationships. The analysed database contains information on the average calories from 16 food aggregates in 40 European countries observed over 40 years by the Food and Agriculture Organisation of the United Nations (FAO). Our findings suggest that European country diets are changing, individually or as a group, but not in a unique direction. Central and Eastern European countries are becoming unhealthier, while the tendency followed by the majority of the remaining countries is to integrate the common European guidelines on healthy, balanced, and diversified diets in their national policies. [ABSTRACT FROM AUTHOR]

By combining linear discriminant analysis and Kmeans into a coherent framework, a dimension reduction algorithm was recently proposed to select the most discriminative subspace. This algorithm utilized the clustering method to generate cluster labels and after that employed discriminant analysis to do subspace selection. However, we found that this algorithm only considers the information of global structure, and does not take into account the information of local structure. In order to overcome the shortcoming mentioned above, this paper presents a dimension reduction algorithm preserving both global and local clustering structure. Our algorithm is an unsupervised linear dimension reduction algorithm suitable for the data with cloud distribution. In the proposed algorithm, the Kmeans clustering method is adopted to generate the clustering labels for all data in the original space. And then, the obtained clustering labels are utilized to describe the global and local clustering structure. Finally, the objective function is established to preserve both the local and global clustering structure. By solving this objective function, the projection matrix and the corresponding subspace are yielded. In this way, the global and local information of the clustering structure are integrated into the process of the subspace selection, in fact, the structure discovery and the subspace selection are performed simultaneously in our algorithm. Encouraging experimental results are achieved on the artificial dataset, real-life benchmark dataset and AR face dataset. [ABSTRACT FROM AUTHOR]

Abstract: Centroids are key components in many data analysis algorithms such as clustering or microaggregation. They are considered as the central value that minimises the distance to all the objects in a dataset or cluster. Methods for centroid construction are mainly devoted to datasets with numerical and categorical attributes, focusing on the numerical and distributional properties of data. Textual attributes, on the contrary, consist of term lists referring to concepts with a specific semantic content (i.e., meaning), which cannot be evaluated by means of classical numerical operators. Hence, the centroid of a dataset with textual attributes should be the term that minimises the semantic distance against the members of the set. Semantically-grounded methods aiming to construct centroids for datasets with textual attributes are scarce and, as it will be discussed in this paper, they are hampered by their limited semantic analysis of data. In this paper, we propose a method that, exploiting the knowledge provided by background ontologies (like WordNet), is able to construct the centroid of multivariate datasets described by means of textual attributes. Special efforts have been put in the minimisation of the semantic distance between the centroid and the input data. As a result, our method is able to provide optimal centroids (i.e., those that minimise the distance to all the objects in the dataset) according to the exploited background ontology and a semantic similarity measure. Our proposal has been evaluated by means of a real dataset consisting on short textual answers provided by visitors of a natural park. Results show that our centroids retain the semantic content of the input data better than related works. [Copyright &y& Elsevier]

Abstract: Web Usage Mining (WUM) is the application of data mining techniques over Web server logs in order to extract navigation usage patterns. The analysis of mining patterns for assessing the knowledge they reveal is a critical phase in WUM. The main challenges are: (a) mining algorithms yield a huge number of patterns and (b) there is a significant semantic gap between URLs and events performed by users. In this paper, we describe the pattern analysis mechanisms integrated in O3R (Ontology-based Rules Retrieval and Rummaging), a human-centered environment for the analysis of navigation rules. O3R explores the synergy of mechanisms for retrieving and analyzing patterns. Filtering and clustering allow users to retrieve subsets of patterns with specific characteristics, in order to deal with the large volume of patterns. Rummaging mechanisms are targeted at assessing the meaning and relevance of pattern with regard to the domain, and it is particularly suitable for exploratory analysis. The distinctive feature of O3R is that is dynamically associates meaning to patterns using the concepts and relationships of a domain ontology, as a means of reducing the gap between syntactic URLs and semantic events performed by users. The paper describes the mechanisms in detail, and explores their synergic integration in the O3R prototype. It also reports two case studies that evaluate the use of O3R for the analysis of navigation patterns of a learning site. [Copyright &y& Elsevier]

Since real-world networks evolve over time, detecting communities in dynamic networks is a challenging research problem with wide applications. In this paper, we first improve our previous method and propose a more stable algorithm which is label-propagation-based for the discovery of communities in complex networks. Then, we present a novel evolutionary clustering approach DLPAE for dynamic networks based on the stable algorithm. According to DLPAE, community labels of nodes are determined by their neighbors, and a confidence (i.e., the importance of its neighbor to the node) is attached to each neighbor. During clustering, the confidences of nodes are calculated in terms of the structures of the current network and the network at last timestamp. We compute confidences’ variance of each node and update nodes’ labels in a descending order according to the values. In our setting, each node can keep one or more labels with belonging coefficients no less than a threshold, which renders DLPAE suitable for detecting overlapping and non-overlapping communities in dynamic networks. Experimental results on both real and synthetic datasets show the ability of DLPAE to detect overlapping and non-overlapping communities in dynamic networks, and demonstrate its higher accuracy compared to other related methods. [ABSTRACT FROM AUTHOR]

Pattern-based clustering algorithms return a set of patterns that describe the objects of each cluster. The most recent algorithms proposed in this approach extract patterns on numerical datasets by applying an a priori discretization process, which may cause information loss. In this paper, we introduce a new pattern-based clustering algorithm for numerical datasets, which does not need an a priori discretization on numerical features. The new algorithm extracts, from a collection of trees generated through a new induction procedure, a small subset of patterns useful for clustering. Experimental results show that the patterns extracted by the proposed algorithm allows to build a pattern-based clustering algorithm, which obtains better clustering results than recent pattern-based clustering algorithms. In addition, the proposed algorithm obtains similar clustering results, in quality, than traditional clustering algorithms. [ABSTRACT FROM AUTHOR]

Abstract: Clustering on categorical data streams is a relatively new field that has not received as much attention as static data and numerical data streams. One of the main difficulties in categorical data analysis is lacking in an appropriate way to define the similarity or dissimilarity measure on data. In this paper, we propose three dissimilarity measures: a point-cluster dissimilarity measure (based on incremental entropy), a cluster–cluster dissimilarity measure (based on incremental entropy) and a dissimilarity measure between two cluster distributions (based on sample standard deviation). We then propose an integrated framework for clustering categorical data streams with three algorithms: Minimal Dissimilarity Data Labeling (MDDL), Concept Drift Detection (CDD) and Cluster Evolving Analysis (CEA). We also make comparisons with other algorithms on several data streams synthesized from real data sets. Experiments show that the proposed algorithms are more effective in generating clustering results and detecting concept drift. [Copyright &y& Elsevier]

Abstract: The problem of finding time series discord has attracted much attention recently due to its numerous applications and several algorithms have been suggested. However, most of them suffer from high computation cost and cannot satisfy the requirement of real applications. In this paper, we propose a novel discord discovery algorithm BitClusterDiscord which is based on bit representation clustering. Firstly, we use PAA (Piecewise Aggregate Approximation) bit serialization to segment time series, so as to capture the main variation characteristic of time series and avoid the influence of noise. Secondly, we present an improved K-Medoids clustering algorithm to merge several patterns with similar variation behaviors into a common cluster. Finally, based on bit representation clustering, we design two pruning strategies and propose an effective algorithm for time series discord discovery. Extensive experiments have demonstrated that the proposed approach can not only effectively find discord of time series, but also greatly improve the computational efficiency. [Copyright &y& Elsevier]

Abstract: Bagging and boosting are two well-known methods of developing classifier ensembles. It is generally agreed that the clusterer ensemble methods that utilize the boosting concept can create clusterings with quality and robustness improvements. In this paper, we introduce a new boosting based hierarchical clusterer ensemble method called Bob-Hic. This method is utilized to create a consensus hierarchical clustering (h-clustering) on a dataset, which is helpful to improve the clustering accuracy. Bob-Hic includes several boosting iterations. In each iteration, first, a weighted random sampling is performed on the original dataset. An individual h-clustering is then created on the selected samples. At the end of the iterations, the individual clusterings are combined to a final consensus h-clustering. The intermediate structures used in the combination are distance descriptor matrices which correspond to individual h-clustering results. This final integration is done through an information theoretic approach. Experiments on popular synthetic and real datasets confirm that the proposed method improves the results of simple clustering algorithms. In addition, our experimental results confirm that this method provides better consensus clustering quality compared to other available ensemble techniques. [Copyright &y& Elsevier]

Abstract: This paper presents a Crotch Ensemble classification model for high dimensional data. A Crotch Ensemble is obtained from a decision cluster tree built by calling a clustering algorithm recursively. A crotch is an inner node of the tree together with its direct children. If the children of a crotch have more than one dominant class, the crotch is defined as a crotch predictor. Each crotch predictor constructs a classifier by itself. A Crotch Ensemble consists of a set of crotch predictors. When classifying a new object, a subset of crotch predictors is selected according to the distances between the object and the crotch predictors. A classification is made on the object as the class predicted by the crotch predictors with the maximum accumulative weights. The experimental results on both synthetic and real data have shown that the Crotch Ensemble model can get better classification results on high dimensional data than other classification methods. [Copyright &y& Elsevier]

Abstract: Clustering analysis is the process of dividing a set of objects into none-overlapping subsets. Each subset is a cluster, such that objects in the cluster are similar to one another and dissimilar to the objects in the other clusters. Most of the algorithms in partitioning approach of clustering suffer from trapping in local optimum and the sensitivity to initialization and outliers. In this paper, we introduce a novel partitioning algorithm that its initialization does not lead the algorithm to local optimum and can find all the Gaussian-shaped clusters if it has the right number of them. In this algorithm, the similarity between pairs of objects are computed once and updating the medoids in each iteration costs O(k × m) where k is the number of clusters and m is the number of objects needed to update medoids of the clusters. Comparison between our algorithm and two other partitioning algorithms is performed by using four well-known external validation measures over seven standard datasets. The results for the larger datasets show the superiority of the proposed algorithm over two other algorithms in terms of speed and accuracy. [Copyright &y& Elsevier]

Abstract: Biclustering is a very useful tool for analyzing microarray data. It aims to identify maximal groups of genes which are coherent with maximal groups of conditions. In this paper, we propose a biclustering algorithm, called BiMine+, which is able to detect significant biclusters from gene expression data. The proposed algorithm is based on two original features. First, BiMine+ is based on the use of a new tree structure, called Modified Bicluster Enumeration Tree (MBET), on which biclusters are represented by the profile shapes of genes. Second, BiMine+ uses a pruning rule to avoid both trivial biclusters and combinatorial explosion of the search tree. The performance of BiMine+ is assessed on both synthetic and real DNA microarray datasets. Experimental results show that BiMine+ competes favorably with several state-of-the-art biclustering algorithms and is able to extract functionally enriched and biologically relevant biclusters. [Copyright &y& Elsevier]

Abstract: Clustering is one of the most important data mining techniques that partitions data according to some similarity criterion. The problems of clustering categorical data have attracted much attention from the data mining research community recently. As the extension of the k-Means algorithm, the k-Modes algorithm has been widely applied to categorical data clustering by replacing means with modes. In this paper, the limitations of the simple matching dissimilarity measure and Ng’s dissimilarity measure are analyzed using some illustrative examples. Based on the idea of biological and genetic taxonomy and rough membership function, a new dissimilarity measure for the k-Modes algorithm is defined. A distinct characteristic of the new dissimilarity measure is to take account of the distribution of attribute values on the whole universe. A convergence study and time complexity of the k-Modes algorithm based on new dissimilarity measure indicates that it can be effectively used for large data sets. The results of comparative experiments on synthetic data sets and five real data sets from UCI show the effectiveness of the new dissimilarity measure, especially on data sets with biological and genetic taxonomy information. [Copyright &y& Elsevier]

Abstract: Semi-supervised learning techniques, such as co-training paradigms, are proposed to deal with data sets with only a few labeled examples. However, the family of co-training paradigms, such as Tri-training and Co-Forest, is likely to mislabel an unlabeled example, thus downgrading the final performance. In practical applications, the labeling process is not always free of error due to subjective reasons. Even some mislabeled examples exist in the few labeled examples given. Supervised clustering provides many benefits in data mining research, but it is generally ineffective with only a few labeled examples. In this paper, a Classification algorithm based on Local Cluster Centers (CLCC) for data sets with a few labeled training data, is proposed. This can reduce the interference of mislabeled data, including those provided by both domain experts and co-training paradigm algorithms. The experimental results on UCI data sets show that CLCC achieves competitive classification accuracy as compared to other traditional and state-of-the-art algorithms, such as SMO, AdaBoost, RandomTree, RandomForest, and Co-Forest. [Copyright &y& Elsevier]

Abstract: In this paper, we introduce a new category of fuzzy models called a fuzzy ensemble of parallel polynomial neural network (FEP2N2), which consist of a series of polynomial neural networks weighted by activation levels of information granules formed with the use of fuzzy clustering. The two underlying design mechanisms of the proposed networks rely on information granules resulting from the use of fuzzy C-means clustering (FCM) and take advantage of polynomial neural networks (PNNs). The resulting model comes in the form of parallel polynomial neural networks. In the design procedure, in order to estimate the optimal values of the coefficients of polynomial neural networks we use a weighted least square estimation algorithm. We incorporate various types of structures as the consequent part of the fuzzy model when using the learning algorithm. Among the diverse structures being available, we consider polynomial neural networks, which exhibit highly nonlinear characteristics when being viewed as local learning models. We use FCM to form information granules and to overcome the high dimensionality problem. We adopt PNNs to find the optimal local models, which can describe the relationship between the input variables and output variable within some local region of the input space. We show that the generalization capabilities as well as the approximation abilities of the proposed model are improved as a result of using polynomial neural networks. The performance of the network is quantified through experimentation in which we use a number of benchmarks already exploited within the realm of fuzzy or neurofuzzy modeling. [Copyright &y& Elsevier]

Abstract: Identification of meaningful clusters from categorical data is one key problem in data mining. Recently, Average Normalized Mutual Information (ANMI) has been used to define categorical data clustering as an optimization problem. To find globally optimal or near-optimal partition determined by ANMI, a genetic clustering algorithm (G-ANMI) is proposed in this paper. Experimental results show that G-ANMI is superior or comparable to existing algorithms for clustering categorical data in terms of clustering accuracy. [Copyright &y& Elsevier]

Community discovery is an important topic of network representation learning. Manifold learning has been widely applied to network representation learning. However, most manifold learning algorithms do not consider the asymmetry of edges which is not accord with the structure of social networks because the influence of nodes is not symmetrical. In this paper, a community discovery algorithm based on manifold graph embedding with structure information propagation mechanism is proposed. The proposed algorithm uses high order approximation matrix to obtain the local and global structure information of a graph, then low rank decomposition is introduced to obtain the node vectors and the context vectors. Finally, the node vectors can be adjusted by structure information. The proposed algorithm and comparison algorithms are conducted on the experimental data sets. The experimental results show that the proposed algorithm outperforms the comparison algorithms on the most experimental data sets. The experimental results prove that the proposed algorithm is an effective algorithm for community discovery. [ABSTRACT FROM AUTHOR]