This paper addresses unsupervised hierarchical classication of personal documents tagged with time and geolocation stamps. The target application is browsing among these documents. A rst partition of the data is built, based on geo-temporal measurement. The events found are then grouped according to geolocation. This is carried out through tting a two-level hierarchy of mixture models to the data. Both mixtures are estimated in a Bayesian setting, with a variational proce- dure: the classical VBEM algorithm is applied for the ner level, while a new variational-Bayes-EM algorithm is introduced to search for suitable groups of mixture components from the ner level. Experimental results are reported on articial and real data.