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.