Exploitation d'une mesure de micro-précision cumulée non supervisée pour l'évaluation fiable de la qualité des résultats de clustering

Abstract : In the context of unsupervised classification, or clustering, the fact of not having a reference classification represents a heavy handicap to evaluate the performance of the algorithms. On their own side, traditional quality indexes (Inertia, DB...) do not allow to properly estimate the quality of the clustering in several cases, as in that one of the textual data. We thus present an alternative approach for clustering quality evaluation based on unsupervised measures of Recall, Precision and F-measure exploiting the descriptors of the data associated with the obtained clusters. The Recall makes it possible to measure the exhaustiveness of the contents of the clusters in terms of peculiar descriptors specific to each cluster. The Precision measures the homogeneity of the clusters in terms of proportion of the data containing the associated peculiar descriptors. This paper especially focuses on the construction of a new cumulative Micro precision index that makes it possible to evaluate the overall quality of a clustering result while clearly distinguishing between homogeneous and heterogeneous results. The experimental comparison of the behavior of the classical indexes with our new index is performed on a dataset of bibliographical references issued from the PASCAL database.