CoConut: Co-Classification with Output Space Regularization

In this work we introduce a new approach to co-classification, i.e. the task of jointly classifying multiple, otherwise independent, data samples. The method we present, named CoConut, is based on the idea of adding a regularizer in the label space to encode certain priors on the resulting labelings. A regularizer that encourages labelings that are smooth across the test set, for instance, can be seen as a test-time variant of the cluster assumption, which has been proven useful at training time in semi-supervised learning. A regularizer that introduces a preference for certain class proportions can be regarded as a prior distribution on the class labels. CoConut can build on existing classifiers without making any assumptions on how they were obtained and without the need to re-train them. The use of a regularizer adds a new level of flexibility. It allows the integration of potentially new information at test time, even in other modalities than what the classifiers were trained on. We evaluate our framework on six datasets, reporting a clear performance gain in classification accuracy compared to the standard classification setup that predicts labels for each test sample separately.