Which Node Attribute Prediction Task Are We Solving? Within-Network, Across-Network, or Across-Layer Tasks

Abstract

Node attribute prediction tasks arise in a wide range of classification tasks on social networks. Examples include detecting spam accounts, identifying compromised accounts, and inferring user demographics for targeted marketing. Despite the prevalence of these types of tasks in machine learning and social science settings, clear problem definitions are lacking. Do all nodes have to be connected in a single network instance? What if there are labels in one network but not another? In this work, we propose a taxonomy that distinguishes between different node attribute prediction tasks; we formalize the existing distinction between within-network and across-network attribute prediction, which have been informally described in prior work, and also introduce a variation we call across-layer attribute prediction. With this framework in place, we observe that methods framed as applicable to across-network tasks have a history of being evaluated on across-layer problem instances. While the methods do well in the across-layer setting, we find that when evaluated in genuine across-network settings, performance can be more limited than previously suggested. We provide a way to analyze and possibly reconcile this predictive performance gap, and highlight why across-network prediction remains an important and open problem domain.

Node Attribute Prediction with Weighted and Directed Edges on Single and Multilayer Networks

“…The typical withinnetwork setting includes known ground-truth labels for a subset of users and a single network on how all users are connected. The goal is then to infer the missing labels or attributes for the remaining nodes (Neville and Jensen 2000;Welling and Kipf 2016;Perozzi, Al-Rfou, and Skiena 2014;Sen et al 2008;Altenburger and Ugander 2021). The related across-layer task trains on a set of labeled nodes and edge type and predicts for the same node set for a different edge type.…”

“…Feature representations for node attribute prediction problems can generally be classified by whether the features are identity-independent (dependent) and label-independent (dependent) (Altenburger and Ugander 2021). Identitydependence captures whether a feature depends on the identity labels of the node i=1,2,...,n. Label-dependence (Gallagher and Eliassi-Rad 2008) captures whether the feature depends on the attribute label.…”

“…While many within-network tasks train on a single network or single edge relationship (Henderson et al 2012(Henderson et al , 2011Altenburger and Ugander 2021;Lim et al 2021), realworld node prediction settings have numerous user-to-user networks to consider for a given prediction task. It is not clear which is the optimal network to select for a given problem.…”