Node Classification with Bounded Error Rates

Abstract

Node classification algorithms are widely used for the task of node label prediction in partially labeled graph data. In many problems, a user may wish to associate a confidence level with a prediction such that the error in the prediction is guaranteed. We propose adopting the Conformal Prediction framework [17] to obtain guaranteed error bounds in node classification problem. We show how this framework can be applied to (1) obtain predictions with guaranteed error bounds, and (2) improve the accuracy of the prediction algorithms. Our experimental results show that the Conformal Prediction framework can provide up to a 30% improvement in node classification algorithm accuracy while maintaining guaranteed error bounds on predictions.