We study in this presentation temporal interaction data: in many situations, directed interactions between actors can be recorded with some reasonably accurate time stamps. This is the case
for instance with email/phone/forum/etc. interactions where the emission instant is recorded by the system under use (email server, cell phone network, etc.). A minimalist representation of such data involves three variables: the emitter, the receiver and the time at which the interaction happened (or started). We propose to explore such data by resorting on a tri-clustering approach in which the values taken on the three dimensions are clustered into classes of emitters, receivers and time stamps that exhibit regularity when considered in conjunction. The proposed model is adjusted to real world data in a fully automated way with a maximum a posteriori approach that needs no user tunable parameters. A post processing technique can be used in addition to over cluster the data when the detailed clusters are too refined. We will illustrate the interested of the proposed approach on real world data.