Topology-Imbalance Learning for Semi-Supervised Node Classification
Dec 6, 2021
About
The class imbalance problem, as an important issue in learning node representations, has drawn increasing attention from the community. Although the imbalance considered by existing studies roots from the unequal quantity of labeled examples in different classes (quantity-imbalance), we argue that graph data expose a unique source of imbalance from the asymmetric topological properties of the labeled nodes, i.e., labeled nodes are not equal in terms of their structural role in the graph (topology imbalance). We propose to unify the two pervasive yet challenging imbalance problems by considering node influence distribution with the Label Propagation algorithm: labeled nodes exert influence of their labeled classes through the graph structure and the imbalance essentially distorts the spread of the influence, resulting in deviation of the model classification boundaries from the true class boundaries. In light of this analysis, we further propose to locate the topological position of the labeled nodes based on influence conflicts from different classes across the graph and devise a model-agnostic method ReNode to re-weight the influence of labeled nodes to address the topology-imbalance problem. Systematic experiments demonstrate the proposed method effectively promotes the performance of various graph neural networks (GNNs) under both the topology- and quantity-imbalance scenarios. Further analysis unveils varied sensitivity of different GNNs to topology imbalance, which may serve as a new perspective in evaluating GNN architectures.The class imbalance problem, as an important issue in learning node representations, has drawn increasing attention from the community. Although the imbalance considered by existing studies roots from the unequal quantity of labeled examples in different classes (quantity-imbalance), we argue that graph data expose a unique source of imbalance from the asymmetric topological properties of the labeled nodes, i.e., labeled nodes are not equal in terms of their structural role in the graph (topolog…
Organizer
NeurIPS 2021
Account · 1.9k followers
About NeurIPS 2021
Neural Information Processing Systems (NeurIPS) is a multi-track machine learning and computational neuroscience conference that includes invited talks, demonstrations, symposia and oral and poster presentations of refereed papers. Following the conference, there are workshops which provide a less formal setting.
Like the format? Trust SlidesLive to capture your next event!
Professional recording and live streaming, delivered globally.
Sharing
Recommended Videos
Presentations on similar topic, category or speaker
Towards Stable and Robust AdderNets
Watch later
Minjing Dong, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Refined Learning Bounds for Kernel and Approximate k-Means
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
AV2.0: Learning a Globally Scalable Driving Intelligence
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Learned Benchmarks for Subseasonal Forecasting
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Generative Models for Hadron Shower Simulation
Watch later
Erik Buhmann, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Evaluation of mathematical questioning strategies using data collected through weak supervision
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%