Section: New Results

TEXT CLASSIFICATION

Neural clustering algorithms show high performance in the general context of the analysis of homogeneous textual datasets. We have recently proposed a new incremental growing neural gas algorithm using the cluster label maximization (IGNGF) [44] [34] . In this strategy the use of a standard distance measure for determining a winner is completely suppressed by considering the label maximization approach as the main winner selection process. One if its important advantage is that it provides the method with an efficient incremental character as it becomes independent of parameters. Although it performs better than the standard clustering methods on textual data, we have shown this year than the obtained results are not as efficient as expected whenever an analysis of very complex heterogeneous textual datasets is performed [33] . We have thus explored several variations of IGNG-F approach based on combination of distance based criteria and cluster label maximization. Our new results on all kinds of datasets, especially on the most complex heterogeneous textual datasets, clearly reflect the advantages of our new algorithm as compared to other existing algorithms and to our former adaptations [29] . Cluster quality evaluation represents a key process for all kinds of data analysis tasks, and more especially for textual data. We have recently presented different variations of unsupervised Recall/Precision and F-measures measures that cope with the defects of classical indexes, like inertia-based indexes. Our new indexes directly exploit the maximized features of the data associated to each cluster after the clustering process without prior consideration of clusters profiles. As compared to classical indexes, their main advantage is thus to be independent of the clustering methods and of their operating mode. They thus altogether permit the objective comparison of clustering methods and represent a sound technique for efficient cluster labeling. We have more especially worked this year on the large scale validation our indexes using reference labeled textual datasets [35] .

We are also currently investigating to set up a platform for efficiently assisting the patents experts in the process of patents validation. Reaching such a goal has implied to develop new semi-supervised classification methods or propose in-deep adaptation of existing ones in order to establish relevant relationships between hierarchical patents classification and bibliographical references describing research covering the fields related to the different patents classes. In this context, we have successfully explored this year new classification techniques based on taboo search [14] .

To cope with the current defects of existing incremental clustering methods, an alternative approach for analyzing information evolving over time consists in performing diachronic analysis. We have thus explored this year different an original technique based on this approach on texts by the use of the combination of cluster labeling with unsupervised Bayesian reasoning between cluster labels extracted from clustering model issued from different time periods. Based on a reference dataset issued from the IST-PROMTECH project, we have clearly shown that these new techniques, whilst providing a new framework for automatizing such kind of analysis, outperformed existing ones [32] [31] [30] .