Dynamic adjusting of neighbourhood sample size with GRAPES
GNNs utilize the concept of aggregating information from neighboring nodes to learn node representations within a graph. As these networks become deeper, their receptive field exponentially increases, leading to significant memory requirements.
To address this, graph sampling techniques have been developed to reduce memory consumption by selecting a subset of nodes from the graph, enabling GNNs to handle larger graphs efficiently. Traditional sampling techniques often rely on static heuristics that might not be effective across various graph structures or tasks. To overcome this limitation, GRAPES was proposed, an adaptive graph sampling strategy that dynamically selects important nodes critical for training GNN classifiers.
GRAPES leverages a GFlowNet to dynamically adjust node sampling probabilities based on the specific goals of classification. To further develop this technique, we propose dynamic adjusting of sample size regulated by a hyperparameter, based on models confidence about certain nodes.