EcoSta 2021: Start Registration
View Submission - EcoSta2021
A0358
Title: Nonparametric link regression and its theoretical properties Authors:  Akifumi Okuno - Institute of Statistical Mathematics (Japan) [presenting]
Keisuke Yano - The University of Tokyo (Japan)
Hidetoshi Shimodaira - Kyoto University and RIKEN AIP (Japan)
Abstract: Given $d$-dimensional data vectors and their link weights, i.e., strengths of graph-structured associations represented by a weighted adjacency matrix, we consider predicting link weights through the corresponding pair of data vectors. We call the problem as link regression (LR), where it reduces to link prediction if the weights are binary. We first apply nonparametric regression methods to the LR setting. Through numerical experiments and theoretical analysis, we also report that their asymptotic behaviors are different depending on the assumed design on the data vectors, i.e., random design and fixed design, unlike the classical regression methods.