Apprentissage non supervisé pour données extrêmement volumineuses en situation de ressources informatiques arbitrairement limitées
Filippo Antonazzo · theses.fr (ABES) · 2022
Clustering reveals all its interest when the data set size considerably increases, since there is the opportunity to discover tiny but possibly high value clusters, which can not be detected with moderate sample sizes. However, the clustering of such high data volumes encounters computational limitations, requiring extremely high memory and computational resources. Thus, current clustering algorithms need frugal implementations, also demanded by institutions and industries to accomplish today's eco-friendly policies. In this context, Gaussian model-based clustering, a popular clustering technique based on Gaussian mixtures, has required frugal adaptations to overcome these computational limitations and to report, even in the huge data case, the same good performance achieved in moderate size analyses. Such implementationsare essentially based on subsampling strategies, which manage to be frugal, but they are expected to heavily failed in highly imbalanced cluster case. Thus, in this work, we propose a frugal technique, based on a so-called bin-marginal data-compression, to perform Gaussian model-based clustering on huge and imbalanced data sets. After a preliminary analysis on simple univariate settings revealing the potential of our solution (here, based on univariate binned data), we extend our proposal to multivariate data sets, where bin-marginal data are employed to perform a drastic reduction of the data volume. Despite this extreme loss of information, we prove identifiability property for the diagonal mixture model and we also introduce a specific EM-like algorithm associated to a composite likelihood approach guaranteeing frugality.Numerical experiments highlight that the proposed method outperforms subsampling both in controlled simulations and in various real applications where imbalanced clusters may typically appear, such as image segmentation, hazardous asteroids recognition and fraud detection. Then, additional topics regarding model choice, the problem of local maxima and the impact of our data-compression on clustering are dealt with a pure experimental point of view. Finally, through a collaboration with a company specialized in predictive maintainance, a practical application of anomaly detection on real time series is shown, in order to extend the potential application domains of the proposal.