Booster is a model-agnostic data-augmentation strategy for temporal knowledge graph completion that addresses imbalanced fact distributions and architecture-specific model preferences. It uses frequency-based filtering plus a hierarchical scoring algorithm built on triadic closures to distinguish false negatives from hard negatives so that generated samples respect the graph's global semantic and local temporal patterns, and a two-stage pre-train/fine-tune procedure to expose models to preference-deviating facts. Students learn how augmentation can be made pattern-aware for dynamic graphs rather than relying on node-connectivity alone.
Temporal knowledge graph completion predicts missing time-stamped facts (subject, relation, object, time) in a dynamic knowledge graph. This paper identifies two neglected obstacles: imbalanced fact …