Summary of Theoretical and Empirical Insights Into the Origins Of Degree Bias in Graph Neural Networks, by Arjun Subramonian et al.
Theoretical and Empirical Insights into the Origins of Degree Bias in Graph Neural Networks
by Arjun Subramonian, Jian Kang, Yizhou Sun
First submitted to arxiv on: 4 Apr 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Social and Information Networks (cs.SI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A new study delves into the phenomenon of Graph Neural Networks (GNNs) favoring high-degree nodes over low-degree nodes in node classification tasks. This degree bias can perpetuate social marginalization by giving more attention to influential individuals or celebrities in online networks. Researchers have proposed various hypotheses for this issue, but a survey of 38 papers reveals that these explanations often lack rigorous validation and may even contradict each other. To address this, the study investigates the origins of degree bias in message-passing GNNs with different graph filters. The results show that high-degree test nodes tend to be misclassified less frequently regardless of training methods, while low-degree nodes are more prone to misclassification. Factors such as homophily and diversity of neighbors contribute to this bias. Furthermore, some GNNs may adjust their loss function more slowly for low-degree nodes during training, but with sufficient epochs, message-passing GNNs can achieve maximum training accuracy. The study connects its findings to previous hypotheses, supporting some while raising doubts about others. Empirical validation is conducted on 8 real-world networks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary GNNs are AI models that perform well for influential people in social media and online networks. But this performance bias can lead to unfairness by prioritizing celebrities over ordinary people. Researchers have tried to explain why GNNs do this, but their theories haven’t been properly tested or agree with each other. To solve this problem, the study looks at how different types of GNNs work on social networks. It shows that high-degree nodes are more accurate and low-degree nodes are less accurate, regardless of how the GNN is trained. This bias comes from factors like who a person’s friends are or how diverse their connections are. The study also finds that some GNNs may take longer to adjust to low-degree nodes during training, but with enough practice, they can do as well as possible. |
Keywords
» Artificial intelligence » Attention » Classification » Gnn » Loss function
