Prompt-Based Zero- and Few-Shot Node Classification: A Multimodal Approach

Multimodal data empowers machine learning models to better understand the world from various perspectives. In this work, we study the combination of \emph{text and graph} modalities, a challenging but understudied combination which is prevalent across multiple settings including citation networks, social media, and the web. We focus on the popular task of node classification using limited labels; in particular, under the zero- and few-shot scenarios. In contrast to the standard pipeline which feeds standard precomputed (e.g., bag-of-words) text features into a graph neural network, we propose \textbf{T}ext-\textbf{A}nd-\textbf{G}raph (TAG) learning, a more deeply multimodal approach that integrates the raw texts and graph topology into the model design, and can effectively learn from limited supervised signals without any meta-learning procedure. TAG is a two-stage model with (1) a prompt- and graph-based module which generates prior logits that can be directly used for zero-shot node classification, and (2) a trainable module that further calibrates these prior logits in a few-shot manner. Experiments on two node classification datasets show that TAG outperforms all the baselines by a large margin in both zero- and few-shot settings.