The problems of link prediction and recovery have been the focus of much work during the last 10 years. This is due to the fact that these questions have a large number of practical implications ranging from detecting spam emails, to predicting which item is selected by which user in a recommendation system. However, considering the highly dynamical aspect of complex networks, there is a rising interest not only for knowing who will interact with whom, but also when. For example, when trying to control the spreading of a virus in a population, it is important to know whether an individual is bound to have a lot of new contacts before or after being infected. In that sense, this question is located at the crossroad of link prediction and another family of problems which has been widely dealt with in the literature, that is, time-series prediction. We name it the interaction prediction problem in link streams. It calls for the definition of specific features, strategies, and evaluation methods to capture both the structural and temporal aspects of the interactions. In this chapter, we propose a general formulation of the problem, consistent with the link stream formalism, which formally represents the streaming sequence of interactions between the elements of the system. Using this framework, we discuss the formulation of the interaction prediction problem and propose possible strategies to address it.