A0169
Title: Conformalized link prediction with FDR control
Authors: Yuan Zhang - The Ohio State University (United States) [presenting]
Abstract: Predicting multiple missing links in partially observed networks is vital across diverse fields. However, controlling the false discovery rate (FDR) in such predictions is challenging due to complex dependencies among test statistics. A conformalized link prediction $({\tt clp})$ method is introduced based on the graphon model that effectively controls FDR, accommodates inhomogeneous missing patterns, and handles unknown missing mechanisms. ${\tt clp}$ constructs conformal p-values row-wise by leveraging the graphon model's exchangeability property and employing a kernel smoothing method to enhance link prediction. Utilizing e-values, it is demonstrated that the Benjamini-Hochberg (BH) procedure successfully controls FDR. Compared to existing methods, {\tt clp} accommodates both binary and weighted networks, provides theoretical guarantees for FDR control, and does not require knowledge of the missing mechanism. The advantages of ${\tt clp}$ are further validated through simulations and real data studies.
