Estimation robuste des modèles de mélange sur des données distribuées

This work proposes a contribution aiming at probabilistic model estimation, in the setting of distributed, decentralized, data-sharing computer systems. Such systems are developing over the internet, and also exist as sensor networks, for instance. Our general goal consists in estimating a probability distribution over a data set which is distributed into subsets located on the nodes of a distributed system. More precisely, we are at estimating the global distribution by aggregating local distributions, estimated on these local subsets. Our proposal exploits the following assumption: all distributions are modelled as a Gaussian mixture. Our contribution is a solution that is both decentralized and statistically robust to outlier local Gaussian mixture models. The proposed process only requires mixture parameters, rather than original data.