Link Prediction with Positive-Unlabeled Learning for Incomplete Graphs

The patent applies preliminary action by constructing an expected graph that includes both observed edges and unobserved missing edges before training the link prediction model. This expected graph is built using domain knowledge and heuristics to predict potential connections that should exist but are currently unobserved. By preparing this augmented graph structure in advance, the model can learn from both connected and unconnected node pairs, preventing overfitting and improving generalization to real-world scenarios where edges may be missing.

The patent applies segmentation by dividing the training data into three distinct segments: observed edges (positive samples), unobserved missing edges (negative samples), and unconnected node pairs (unlabeled samples). Each segment is processed differently during training - observed edges provide confirmed connection patterns, unobserved edges provide counter-examples of non-connections, and unconnected pairs provide additional unlabeled data for self-supervised learning. This segmented approach allows the model to efficiently utilize all available information while maintaining training productivity through specialized processing for each data type.

The patent introduces an expected graph as an intermediary structure that bridges the gap between the given edge-incomplete graph and the target prediction task. This expected graph incorporates domain knowledge, heuristics, and preliminary predictions to augment the original graph with potential missing edges. The link prediction model is then trained on this enriched expected graph rather than directly on the incomplete original graph, allowing it to learn more accurate connection patterns while maintaining training feasibility. The intermediary expected graph acts as a mediator that transforms the incomplete input into a more informative training representation.