Leveraging Free Labels to Power up Heterophilic Graph Learning in Weakly-Supervised Settings: An Empirical Study

MACHINE LEARNING AND KNOWLEDGE DISCOVERY IN DATABASES: RESEARCH TRACK, ECML PKDD 2023, PT III(2023)

Cited 0|Views9
No score
Abstract
Graph learning on heterophilic graphs is challenging for classic graph neural network models. Recent research addresses this issue by using adaptive graph filters that consider signals with different frequencies. Although such models provide insightful design patterns for heterophilic graph analysis, their practical effect has been overlooked. Previous studies have evaluated adaptive graph filters with a large proportion of training data to demonstrate their effectiveness. However, such dense labeling is often impractical. Empirically, we observed that typical adaptive filters perform badly in weakly-supervised settings, making them easily outperformed by fixed filters. With empirical evidence, we demonstrate that the key reason is that sparse node labels make it difficult to learn effective filters. Fortunately, although labeled nodes are sparse in weakly-supervised settings, graph structures provide substantial supervision by indicating whether an edge is present. Through theoretical analysis on contextual Stochastic Block Models, we show that a good link predictor can imply the knowledge needed by a good node classifier. Therefore, we propose to use the "free labels" from the graph structure to form link prediction tasks and obtain an effective graph filter, which can be used to initialize the node classification model. Experimental results on both synthetic and real-world heterophilic graphs demonstrate the effectiveness of our approach. We also provide an in-depth analysis of the learned filters, which sheds light on the underlying mechanisms of our proposed approach. Codes are available at https://github.com/lucio-win/PKDD2023.
More
Translated text
Key words
Heterophilic Graph Learning,Adaptive Filter,Weakly-Supervised Learning
AI Read Science
Must-Reading Tree
Example
Generate MRT to find the research sequence of this paper
Chat Paper
Summary is being generated by the instructions you defined