A non-aggregative quantification method for graph nodes

Node quantification is the task of estimating the prevalence of di!erent node labels in arbitrary subsets of nodes extracted from graphs, under prior probability shift. Existing approaches to node quan- tification typically follow the aggregative paradigm, i.e., they train an intermediate node classifier, aggregate its predictions, and then correct the resulting prevalence estimates using various adjustment strategies. In this work we propose a non-aggregative method that estimates node prevalence values directly from sets of node representations, thereby by- passing the need for an intermediate node classifier. Notably, these aggre- gated representations are obtained from an untrained Graph Echo State Network (GESN), which ensures e”ciency especially during inference. Our approach achieves performance comparable to aggregative methods while preserving the distinctive advantage of remaining applicable when inferring individual node labels is undesirable due to privacy concerns.