International audience

Clustering is often a first step when trying to make sense of a large data set. A wide family of cluster analysis algorithms, namely hierarchical clustering algorithms, does not provide a partition of the data set but a hierarchy of clusters organized in a binary tree, known as a dendrogram. The dendrogram has a classical node-link representation used by experts for various tasks like: to decide which subtrees are actual clusters (e.g., by cutting the dendrogram at a given depth); to give those clusters a name by inspecting their content; etc. We present Dendrogramix, a hybrid tree-matrix interactive visualization of dendrograms that superimposes the relationship between individual objects on to the hierarchy of clusters. Dendrogramix enables users to do tasks which involve both clusters and individual objects that are impracticable with the classical representation, like: to explain why a particular objects belongs to a particular cluster; to elicit and understand uncommon patterns (e.g., objects that could have been classified in a totally different cluster); etc. Those sensemaking tasks are supported by a consistent set of interaction techniques that facilitates the exploration of large clustering results.

In this paper, we introduce a new Focus+Context visualization technique, named "Stacked Trees", allowing us to explore large dendrograms produced by hierarchical clustering. This approach displays up to fifty thousands nodes on a standard-sized screen.

In many applications, entities of the domain are described through different views that clustering methods often process one by one. We introduce here the architecture MVSim, that is able to deal simultaneously with all the information contained in such multi-view datasets by using several instances of a co-similarity algorithm. We show that this architecture provides better results than both single-view and multi-view approaches and that it can be easily parallelized thus reducing both time and space complexities of the computations.

The classical representation of a binary tree generated by a hierarchical clustering is a node-link-based visualization denoted as a dendrogram. It allows users to explore in a simple way the clusters and the relationships between instances. However, exploration of large dendrograms is known to be difficult due to the graphical and cognitive information overload involved. Here, we discuss the current approaches and we introduce Stacked Trees, a new Focus+Context visualization technique that allows the exploration of the hierarchical clustering of up to fifty thousands nodes on a standard-sized screen.

Co-clustering has been defined as a way to organize simultaneously subsets of instances and subsets of features in order to improve the clustering of both of them. In previous work, we proposed an efficient co-similarity measure allowing to simultaneously compute two similarity matrices between objects and features, each built on the basis of the other. Here we propose a generalization of this approach by introducing a notion of pseudo-norm and a pruning algorithm. Our experiments show that this new algorithm significantly improves the accuracy of the results when using either supervised or unsupervised feature selection data and that it outperforms other algorithms on various corpora.

In this paper, we propose an extension of the χ-Sim co-clustering algorithm to deal with the text categorization task. The idea behind χ-Sim method [1] is to iteratively learn the similarity matrix between documents using similarity matrix between words and vice-versa. Thus, two documents are said to be similar if they share similar (but not necessary identical) words and two words are similar if they occur in similar documents. The algorithm has been shown to work well for unsupervised document clustering. By introducing some “a priori” knowledge about the class labels of documents in the initialization step of χ-Sim, we are able to extend the method to deal for the supervised task. The proposed approach is tested on different classical textual datasets and our experiments show that the proposed algorithm compares favorably or surpass both traditional and state-of-the-art algorithms like k-NN, supervised LSI and SVM.

Co-clustering has been widely studied in recent years. Exploiting the duality between objects and features efficiently helps in better clustering both objects and features. In contrast with current co-clustering algorithms that focus on directly finding some patterns in the data matrix, in this paper we define a (co-)similarity measure, named X-Sim, which iteratively computes the similarity between objects and their features. Thus, it becomes possible to use any clustering methods (k-means, ...) to co-cluster data. The experiments show that our algorithm not only outperforms the classical similarity measure but also outperforms some co-clustering algorithms on the document-clustering task.