Link Prediction

Foundational link prediction relies on topological heuristics that score node pairs based on graph structure without requiring training data.

• Common Neighbors: Counts shared adjacent nodes—two machines sharing many failure modes are likely related
• Jaccard Coefficient: Normalizes common neighbors by total neighborhood size to penalize highly-connected hub nodes
• Adamic-Adar Index: Weighs shared neighbors inversely by their degree, giving more weight to rare, specific connections
• Preferential Attachment: Predicts links based on node degree product, capturing the rich-get-richer phenomenon in supplier networks

These methods provide interpretable baselines for identifying potential causal relationships between equipment failures before deploying complex models.

Modern link prediction uses Graph Neural Networks (GNNs) to learn low-dimensional vector representations that encode both node attributes and local graph topology.

• Graph Convolutional Networks (GCNs) aggregate features from multi-hop neighborhoods, capturing that a vibration anomaly on Pump-A may predict failure on downstream Compressor-B
• GraphSAGE enables inductive learning on previously unseen nodes, critical when new equipment is commissioned into an existing manufacturing line
• Edge features such as material flow rates or thermal transfer coefficients are incorporated directly into the message-passing framework

These embeddings transform the link prediction task into a binary classification problem in embedding space, enabling the discovery of non-obvious, higher-order dependencies.

Manufacturing relationships evolve over time, and link prediction must account for temporal dynamics to avoid stale inferences.

• Exponential decay weighting assigns higher importance to recent co-occurrence events—a failure pattern observed last week is more predictive than one from five years ago
• Hawkes process models capture self-exciting behavior where one failure event temporarily increases the probability of linked failures
• Time-aware negative sampling ensures the model learns to distinguish genuinely absent links from those that simply haven't manifested yet

Temporal knowledge graphs with valid-time intervals on edges allow the system to predict not just if a link exists, but when it is likely to form.

Manufacturing knowledge graphs contain heterogeneous edge types—failure modes, material flows, supply dependencies, and maintenance schedules—each requiring distinct prediction logic.

• DistMult and ComplEx factorization models learn relation-specific transformations that capture asymmetric dependencies like causesFailure versus mitigatedBy
• RotatE models relations as rotations in complex vector space, naturally encoding the compositional logic that if Machine-A feeds Machine-B and Machine-B feeds Machine-C, then Machine-A indirectly supplies Machine-C
• Type-constrained decoding restricts predictions to valid relation-domain-range combinations, preventing nonsensical links like a sensor consuming a work order

This enables simultaneous prediction across failure propagation, supplier qualification, and maintenance scheduling edges from a unified model.

Since knowledge graphs only contain positive assertions, link prediction critically depends on intelligent negative sampling to teach models what connections should not exist.

• Uniform random sampling provides a weak baseline but generates trivially easy negatives
• Corrupting positive triples by replacing head or tail entities creates harder negatives that share partial context
• Adversarial sampling dynamically selects negatives the current model is most confused about, accelerating convergence
• Domain-constrained sampling ensures negatives respect manufacturing physics—a hand tool cannot be a valid substitute for a CNC machine

Effective negative sampling prevents the model from degenerating into predicting all possible links with high confidence.

For manufacturing decision-making, link predictions must be auditable and well-calibrated to justify interventions like preemptive maintenance or supplier changes.

• GNNExplainer identifies the minimal subgraph and node features that drove a specific prediction, revealing that a link was predicted because two machines share a specific lubricant type
• Platt scaling and isotonic regression calibrate raw model scores into well-calibrated probabilities, ensuring a 0.8 confidence score truly reflects 80% likelihood
• Counterfactual explanations show what minimal graph changes would flip the prediction, helping engineers understand boundary conditions
• Conformal prediction provides statistically valid confidence sets with guaranteed coverage probabilities for high-stakes applications

These techniques transform link prediction from a black-box recommendation into a trusted engineering tool.

A class of deep learning models designed to operate directly on graph-structured data. GNNs learn node representations by aggregating feature information from neighboring nodes, making them the primary architecture for link prediction tasks. In manufacturing, GNNs can predict missing connections in a Bill of Materials Graph or classify machine health from a sensor network topology. Common variants include Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs).

The fundamental data structure of a knowledge graph consisting of a subject-predicate-object statement. For example, Pump-23 hasFailureMode BearingFatigue encodes a single fact. Link prediction operates on these triples by scoring the plausibility of unobserved subject-predicate-object combinations, effectively asking: 'Is this missing triple likely to be true?'

A directed acyclic graph (DAG) that encodes cause-and-effect relationships between manufacturing variables. Unlike correlation-based link prediction, causal graphs enable engineers to simulate interventions and answer counterfactual questions. When a link prediction model suggests a new dependency between a temperature spike and a quality defect, a causal graph validates whether the relationship is actionable or spurious.

A knowledge graph that explicitly models the time dimension of facts, allowing queries about system state at any historical point. Link prediction in temporal graphs forecasts future connections—such as which component is likely to fail next in a sequence—by learning from ordered event chains. This is critical for predictive maintenance and failure cascade analysis.

The computational task of disambiguating records that refer to the same real-world asset across disparate data sources. Before link prediction can operate effectively, entity resolution must create a unified golden record—ensuring that 'Pump-23' in the maintenance log is the same node as 'PUMP_023' in the SCADA system. Without this deduplication, link prediction models produce fragmented, unreliable results.

A software component that applies logical inference rules to a knowledge graph's ontology to derive new, implicit facts from explicitly asserted data. While link prediction uses statistical machine learning to suggest probable connections, a reasoner uses formal logic to deduce certain connections. The two techniques are complementary: link prediction proposes hypotheses, and the reasoner validates them against domain constraints like SHACL shapes.