Core Concepts
The author proposes Disentangled Intervention-based Dynamic graph Attention networks with Invariance Promotion (I-DIDA) to handle spatio-temporal distribution shifts in dynamic graphs by discovering and utilizing invariant patterns.
Abstract
The paper introduces I-DIDA, a method to address spatio-temporal distribution shifts in dynamic graphs. It focuses on discovering invariant patterns to improve generalization under distribution shifts. Extensive experiments demonstrate the superiority of I-DIDA over state-of-the-art baselines.
Dynamic graph neural networks have shown powerful predictive abilities but struggle with distribution shifts. The proposed I-DIDA method aims to capture invariant patterns for stable predictions. By leveraging disentangled attention mechanisms and intervention strategies, the model can handle complex structural and temporal information effectively.
The approach involves disentangling spatio-temporal patterns, creating intervened distributions, and inferring latent environments to minimize prediction variance. By focusing on invariant patterns, the model achieves superior performance in link prediction tasks across various datasets.
Stats
Extensive experiments demonstrate the superiority of our method over state-of-the-art baselines under distribution shifts.
The dataset includes 23,035 nodes and 151,790 links in total.
We use "Data Mining" as ‘w/ DS’ and the left as ‘w/o DS’.
We use word2vec [75] to extract 32-dimensional features from paper abstracts.
The dataset includes 13,095 nodes and 65,375 links in total.
We use "Pizza" as ‘w/ DS’ and the left as ‘w/o DS’.
Quotes
"Our work is the first study of spatio-temporal distribution shifts in dynamic graphs."
"We propose Disentangled Intervention-based Dynamic Graph Attention Networks with Invariance Promotion (I-DIDA)."
"The contributions of our work are summarized as follows."