Symmetry-Aware GFlowNets (SA-GFN) corrects biases in graph sampling by integrating symmetry corrections into the reward structure, improving the diversity and accuracy of generated graphs.
Generative Flow Networks (GFlowNets) are used to sample graphs, but they often have biases due to the symmetrical nature of graphs, which affects how accurately they can generate samples proportional to their rewards. The new method, Symmetry-Aware GFlowNets (SA-GFN), addresses this by adjusting the reward system to account for these biases, eliminating the need for complex calculations of state transitions. This approach not only removes the biases but also improves the diversity and quality of the generated graphs, making them more consistent with the desired outcomes.
