FlowNeg Improves Knowledge Graph Embeddings with Diverse Negative Sampling

The effectiveness of Knowledge Graph Embedding (KGE) models heavily relies on the quality of negative samples used during training. Traditional methods often struggle, with uniform negatives being too easy and hard-negative miners tending to focus on a limited set of entities, potentially colliding with actual positive examples. This research presents FlowNeg, a novel approach to address these limitations. FlowNeg utilizes a context-conditioned hierarchical generative flow network (GFlowNet) to amortize reward-proportional sampling. Given a positive triple, FlowNeg intelligently selects a type and then an entity for corruption. Its unique terminal reward function combines a bounded model-based hardness score with a training-only structural score, which helps prevent collisions with held-out positive examples and ensures type compatibility. Extensive experiments across multiple architectures and benchmarks demonstrated FlowNeg's superior performance. It consistently achieved higher mean MRR (Mean Reciprocal Rank) compared to leading hard-negative sampling methods like EMU and IF-NS, with significant average improvements. FlowNeg also maintained high gradient informativeness and low collision rates, supporting its ability to generate diverse, mode-covering negative samples without relying on structural similarity as an absolute truth oracle.