A classification task requires the training dataset to have examples (observations) for all the target classes that will be observed in the test set. Let's just consider a classification task where you would like to extract an animal species from a picture, whatever the animal is. Recent estimates assume that there are around eight million species on the Earth. This means that we will have to build an image dataset with millions of images. In order not to reach these numbers, zero-shot learning tries to infer classes in the test set, even if they are not in the training set.
This is an ongoing research topic and several solutions have been proposed to tackle this problem. One approach using GNNs consists of the following idea. Starting from a knowledge graph where each node is a class, connected depending on some attribute's similarity, a GNN is trained whose task is to output a classifier for each class. In practice, the GNN learns the weights of the output...
