---

name: knowledge-graphs description: "Use TorchDrug for knowledge graph reasoning, triple datasets, embedding models, NeuralLP, KGC tasks, negative sampling, and filtered ranking troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Knowledge Graphs

Use this sub-skill when a user asks about knowledge graph reasoning, link prediction over triples, KGC datasets, filtered ranking metrics, negative sampling, or switching between TorchDrug embedding models and NeuralLP.

When to Use

• The data is a set of (head, relation, tail) facts or benchmark knowledge graph triples.
• The task is knowledge graph completion, missing-link prediction, filtered ranking evaluation, or rule-style reasoning.
• The user mentions FB15k, FB15k-237, WN18, WN18RR, YAGO3-10, Hetionet, TransE, DistMult, ComplEx, RotatE, SimplE, KBGAT, or NeuralLP.
• The user needs task-parameter guidance for KnowledgeGraphCompletion, including num_negative, fact_ratio, sample_weight, strict_negative, filtered_ranking, or full_batch_eval.

Quick Routing

• For generic graph tensors, packing, masking, graph construction, or data.Graph mechanics, use the sibling graph-data sub-skill.
• For core.Engine, optimizers, schedulers, checkpoints, logging, and multi-GPU training loops, use the sibling training-engine sub-skill.
• For KG-specific model/task/dataset decisions, start with references/api-reference.md.
• For end-to-end embedding and NeuralLP recipes, use references/reasoning-workflows.md.
• For bad triples, relation/entity indexing errors, negative sampling surprises, dataset download issues, or ranking-memory failures, use references/troubleshooting.md.

Safe Helper

Run scripts/plan_kg_reasoning.py to print a side-effect-free API skeleton for an embedding or NeuralLP workflow. It does not download datasets, import TorchDrug, allocate GPUs, or train models.