AUC_ConceptFig.pdf

Flowchart describing the application of Area-Under-the-Curve (AUC) metric evaluation of a link-prediction model. Often used for assessing the performance of models of connectivity. Here the model has predicted a specific clustering of the nodes in the adjacency matrix, represented on the left as the coloured edges, where the grey-level shading indicates the estimated link-density of that particular link (connection between clusters). The node-level links from a test set are overlaid as dots. The better models should therefore have cluster-level links that are darker where there are many test node-level links, and lighter where there are fewer. The receiver-operating characteristic (ROC) curve for a particular clustering is generated by sweeping through a range of thresholds for the cluster-level links, and for each one evaluating the number of correctly- and incorrectly-predicted node-level links that exist. These points form an ROC curve, and the area beneath it is the AUC value for this particular clustering.