We propose and analyze two strategies to learn over unordered pairs with kernels, and provide a common theoretical framework to compare them. The strategies are related to methods that were recently investigated to predict edges in biological networks. We show that both strategies differ in their loss function and in the kernels they use. We deduce in particular a smooth interpolation between the two approaches, as well as new ways to learn over unordered pairs. The different approaches are

more »

... ed on the inference of missing edges in two biological networks.