Capturing both structural and temporal features of interactions is crucial in many real-world situations like studies of contact between individuals. Using the link stream formalism to model data, we address here the activity prediction problem: we predict the number of links that will occur during a given time period between each pair of nodes. To do this, we take benefit from the temporal and structural information captured by link streams. We design and implement a modular supervised learning method to make prediction, and we study the key elements influencing its performances. We then introduce classes of node pairs, which improves prediction quality and increases diversity.