Bayesian Graph Convolutional Neural Networks for Noisy Data

The patent applies preliminary action by generating multiple perturbed graph realizations before training the GCNN model. Instead of directly training on the observed graph, the method first creates multiple alternative graph structures by adding/removing edges according to a probabilistic model, then trains the GCNN on these pre-generated realizations. This preliminary perturbation process prepares the model to handle noisy data by exposing it to various possible graph configurations before actual training begins.

The patent implements beforehand cushioning by introducing controlled noise and uncertainty into the graph data before training. By generating multiple graph realizations with varying edge configurations, the method creates a cushion of diversity that protects the model from overfitting to any single graph configuration. This cushioning effect allows the model to learn robust patterns that generalize better to noisy, real-world graph data.

The patent applies dynamics by making the graph structure dynamic rather than static. Instead of aggregating from a fixed neighbor set, the method generates multiple dynamic graph realizations where the neighbor relationships vary across realizations. The GCNN is trained on these dynamic variations, allowing it to learn which neighbor relationships are truly informative versus those that are artifacts of a single graph configuration. This dynamic approach prevents overfitting while preserving structural information.

The patent implements parameter changes by varying the graph structure parameters (edge presence/absence) across multiple realizations. By changing the adjacency matrix parameters according to a probabilistic model, the method creates diverse training samples with different neighbor configurations. This parameter variation forces the model to learn robust features that are invariant to specific graph configurations, thereby preventing overfitting while maintaining structural information capture.

The patent applies partial action by generating a limited number of graph realizations rather than exhaustively sampling all possible graph configurations. Instead of considering every possible edge perturbation, the method generates a manageable set of realizations (e.g., 10-100 samples) that provide sufficient diversity for robust training without overwhelming computational resources. This partial sampling approach balances uncertainty measurement capability with computational feasibility.

The patent implements segmentation by dividing the training process into multiple independent stages, each working on a separate graph realization. Rather than processing one large complex task, the method segments the problem into multiple smaller, parallelizable training runs on individual graph realizations. This segmentation allows the computational workload to be distributed and managed efficiently, reducing overall complexity while maintaining the benefits of multiple realizations.