While deep learning has achieved great success on various tasks, the
task-specific model training notoriously relies on a large volume of labeled
data. Recently, a new training paradigm of ``pre-train, prompt, predict'' has
been proposed to improve model generalization ability with limited labeled
data. The main idea is that, based on a pre-trained model, the prompting
function uses a template to augment input samples with indicative context and
reformalizes the target task to one of the pre-training tasks. In this survey,
we provide a unique review of prompting methods from the graph perspective.
Graph data has served as structured knowledge repositories in various systems
by explicitly modeling the interaction between entities. Compared with
traditional methods, graph prompting functions could induce task-related
context and apply templates with structured knowledge. The pre-trained model is
then adaptively generalized for future samples. In particular, we introduce the
basic concepts of graph prompt learning, organize the existing work of
designing graph prompting functions, and describe their applications and
challenges to a variety of machine learning problems. This survey attempts to
bridge the gap between structured graphs and prompt design to facilitate future
methodology development.

通过基于结构化图形的触发方法，绕过传统的大规模标注数据集，提高深度学习的泛化能力，使预训练模型更具适应性。