The term "dyad ranking" refers to a new problem setting within preference learning. Dyads are feature vector pairs that need to be ranked by machine learning models. Existing ranking methods do not deliver good results for dyad ranking, since they do not use all features of the dyads. Therefore, three generalizations of the Plackett-Luce (PL) model, a statistical model for rank data, are introduced: Joint-Feature PL (JFPL) uses joint-feature vector representations for the dyads, i.e. a mapping of a vector pair to a single vector. The bilinear PL model (BilinPL), which takes up the idea of JFPL, specifies the joint-feature map by means of the cross product. Experiments show that BilinPL is superior to existing label ranking methods, because the dyad features improve prediction performance and it can deliver predictions on new labels. The third model, PLNetworks (PLNet), does not require the specification of a joint-feature map but instead learns it. The model is based on a neural network and can capture non-linear relationships among preferences. Applications of dyad ranking include genetic algorithm recommendations, similarity learning of images, and the configuration of image-processing pipelines using preference- based reinforcement learning. To benefit from the probabilistic information produced by the PL models, two visualization approaches based on multidimensional scaling and unfolding are introduced.