In this paper, we study the problem of source detection in the contextof information diffusion through online social networks. We propose a representationlearning approach that leads to a robust model able to deal with the sparsityof the data. From learned continuous projections of the users, our approachis able to efficiently predict the source of any newly observed diffusion episode.Our model does not rely neither on a known diffusion graph nor on a hypotheticalprobabilistic diffusion law, but directly infers the source from diffusion episodes.It is also less complex than alternative state of the art models. It showed goodperformances on artificial and real-world datasets, compared with various stateof the art baselines.