Knowledge Graph Explanations

Identifies the specific sequences of triples (subject-predicate-object) that connect a source node to a target prediction. These paths serve as human-readable justifications for why a link was predicted.

• Breadth-First Search (BFS) with learned attention weights to score path importance
• Reinforcement learning agents that walk the graph to find reward-maximizing paths
• Example: Explaining a "treats" prediction between a drug and disease by surfacing the path Drug → targets → Protein → associated_with → Disease

Distills a trained knowledge graph embedding model into a set of symbolic, first-order logical rules that approximate its predictions. These Horn clauses provide a fully interpretable surrogate model.

• Rules take the form: (e1, r1, e2) ∧ (e2, r2, e3) ⇒ (e1, rH, e3)
• Confidence scores quantify the statistical support for each rule in the graph
• Example: bornIn(X, Y) ∧ locatedIn(Y, Z) ⇒ nationality(X, Z)

Assigns a numerical saliency score to each triple in the neighborhood of a prediction, quantifying its contribution to the model's decision. This decomposes the output into its most influential facts.

• Gradient-based methods compute the partial derivative of the prediction score with respect to each triple's embedding
• Perturbation analysis measures the drop in prediction confidence when a triple is masked
• Enables fine-grained auditing of which facts the model relied upon

Extracts a compact, connected explanatory subgraph surrounding the predicted link or node. The subgraph contains only the entities and relations most critical to the decision, discarding irrelevant context.

• Mutual information maximization selects subgraphs that preserve predictive information
• Graph Information Bottleneck principles compress the input to a minimal sufficient statistic
• Provides a visual, structural justification analogous to a citation network for the prediction

Identifies the minimal set of triple additions or deletions that would flip the model's prediction. This reveals the decision boundary and provides actionable recourse.

• Answers: "What missing relationship would have changed this classification?"
• Uses structural causal models to distinguish correlational from causal relationships
• Critical for debugging erroneous predictions and ensuring model robustness against graph perturbations

Decomposes the learned entity and relation embeddings into interpretable dimensions corresponding to human-understandable concepts. This bridges the gap between latent representations and symbolic knowledge.

• Tensor decomposition techniques like CP or Tucker factorize embeddings into concept-specific components
• Disentangled representation learning enforces independence between explanatory factors
• Enables queries like: "Which dimensions of this drug embedding encode its mechanism of action?"