Temporal Link Prediction

The core assumption is that future graph structure is dependent on past structure. Models must capture sequential patterns, recurrence, and causality in edge formation. This is distinct from static models that treat all observed links as contemporaneous.

• Example: Predicting a future co-authorship between two researchers based on their evolving publication history and collaboration networks over the past decade.

Entities and relations are embedded in a vector space where their representations evolve over time. Temporal Knowledge Graph Embeddings (TKGE) like TTransE, DE-SimplE, or TeRo learn functions where the score for a triple (subject, relation, object) depends on a timestamp.

• Key Technique: Encoding time via temporal encoding functions (e.g., sinusoidal, continuous time) or by treating time as an additional dimension in the embedding space.

Predictions are made for specific future time points or intervals. The model's output is conditioned on a query time t_q. Data can be represented as a sequence of graph snapshots (discrete time) or as a stream of timestamped events (continuous time).

• Granularity Choices: Predictions can be for the next day, quarter, or year, depending on the application domain (e.g., social networks vs. financial transactions).

Models are built using specialized neural architectures designed for sequential, structured data:

• Recurrent Architectures: Use RNNs or LSTMs to process sequences of graph snapshots.
• Temporal Graph Neural Networks (TGNNs): Extend GNNs with mechanisms to aggregate information from a node's temporal neighborhood (e.g., TGCN, EvolveGCN).
• Attention-Based Models: Use self-attention (like in Transformers) to weigh the importance of past events when predicting a future link.

Performance is measured using time-sensitive evaluation protocols. The dataset is split chronologically into training, validation, and test sets. The model is trained on past data (t < t_train) and evaluated on its ability to predict links in future periods (t >= t_test).

• Primary Metrics: Time-sensitive variants of Mean Reciprocal Rank (MRR) and Hits@k are standard, assessing the model's ranking of true future links against corrupted negatives.

This capability powers predictive systems in dynamic, networked environments:

• Recommendation Systems: Predicting future user-item interactions in e-commerce or media.
• Anomaly Detection: Forecasting unusual financial transactions or network intrusions.
• Scientific Discovery: Anticipating new research collaborations or drug-target interactions.
• Supply Chain Risk: Predicting future disruptions in logistic or supplier networks.

Predicts future supplier-customer or warehouse-transport links to optimize inventory and routing.

• Dynamic Routing: Forecasts carrier availability and port congestion to preemptively reroute shipments.
• Supplier Risk: Anticipates the dissolution of links with at-risk vendors, enabling proactive sourcing.
• Demand Forecasting: Models future product-retailer connections based on seasonal and promotional trends.

Forecasts illicit transaction networks (e.g., money laundering) by modeling the temporal evolution of account interactions.

• Transaction Graph Evolution: Models how fraudulent networks slowly build trust (link formation) before executing large transfers.
• Temporal Anomaly Detection: Flags link formations that deviate from an entity's historical behavioral pattern.
• Synthetic Identity Fraud: Predicts future links between synthetic identities and real accounts based on grooming patterns.

Anticipates the formation of future social or professional connections, informing recommendation systems and churn prediction.

• Professional Networks (e.g., LinkedIn): Suggests future collaborations or mentor-mentee relationships based on evolving project involvement and skill endorsements.
• Churn Prediction: Models the weakening (potential dissolution) of user-community links to trigger retention interventions.
• Influence Mapping: Forecasts the rise of new key opinion leaders by predicting rapid future link formations around emerging topics.

Predicts future communication paths between compromised devices (lateral movement) in an enterprise network.

• Attack Graph Forecasting: Models how an attacker might propagate from an initial breach point to critical assets based on historical vulnerability patching cycles and access logs.
• Botnet Recruitment: Anticipates which vulnerable IoT devices are likely to be linked into a botnet based on scanning activity patterns.
• Insider Threat: Identifies unusual future link formations between an employee and sensitive data repositories outside normal patterns.

Models the future spread of diseases through contact networks or predicts patient-care provider interactions.

• Epidemic Forecasting: Predicts future transmission links in a population contact network to target vaccination campaigns.
• Patient Readmission: Forecasts future patient-facility links (readmissions) based on temporal care pathways and treatment histories.
• Clinical Trial Recruitment: Identifies future suitable patient-investigator links by modeling eligibility criteria and patient health state evolution.

Forecasts a user's future interaction with items (products, content, services) by modeling the user-item graph as a dynamic network.

• Sequential Recommendation: Predicts the next item in a session based on the temporal pattern of past user-item links.
• Lifecycle Marketing: Anticipates when a user is likely to form a link with a premium service tier based on their usage evolution.
• Market Basket Evolution: Models how product co-purchase networks change over seasons, predicting future complementary item links.

A knowledge graph that explicitly represents the time-varying nature of facts, entity states, and relationships by associating them with temporal validity intervals or timestamps. This is the foundational data structure for temporal link prediction.

• Core Components: Nodes, edges, and properties are all annotated with time metadata.
• Representation: Often modeled as a sequence of static graph snapshots or as a stream of timestamped events.
• Use Case: Enables queries like "What suppliers did this manufacturer work with in Q3 2023?"

A technique that learns low-dimensional vector representations for entities and relations in a temporal knowledge graph, capturing both semantic meaning and temporal relational patterns. These embeddings are the primary input features for many temporal link prediction models.

• Goal: Encode how relationships evolve (e.g., a 'worksAt' relation may change every few years).
• Methods: Include models like TTransE, DE-SimplE, and ATiSE, which incorporate time as a transformation or separate embedding.
• Output: A time-aware vector for each entity that changes or is contextualized by time.

A class of neural network architectures designed to learn representations from dynamic graph data by incorporating temporal dependencies into the message-passing or aggregation process. TGNNs are state-of-the-art models for temporal link prediction.

• Mechanism: They aggregate information from a node's neighbors across recent graph snapshots.
• Architectures: Include Temporal Graph Convolutional Networks (TGCN) and attention-based models like DySAT.
• Advantage: Can capture complex, non-linear temporal evolution patterns that simpler embedding models miss.

A temporal knowledge graph model centered on events as first-class entities, with relationships capturing temporal, causal, and participative links between events and entities. Link prediction in event graphs often forecasts future events or participations.

• Structure: Nodes represent events (e.g., 'Transaction_459', 'System_Alert') and entities (e.g., 'Customer_A', 'Server_05').
• Relations: Include 'causes', 'precedes', 'involves', and 'hasParticipant'.
• Application: Ideal for modeling business processes, IT incident chains, or financial transaction networks.

A graph whose structure (nodes and edges) changes over time, serving as a general mathematical model for temporal and streaming graph data. Temporal link prediction is a key analytical task on dynamic graphs.

• Scope: Broader than TKGs; may not have formal semantic ontologies.
• Representation: Often modeled as a graph stream G(t) = (V(t), E(t)).
• Analysis Tasks: Include forecasting, anomaly detection, and community tracking, all relying on understanding link dynamics.

The broader task of inferring missing facts in a temporal knowledge graph, where predictions must be accurate for a specific query time or validity interval. Temporal link prediction is a core sub-task of TKGC.

• Objective: Predict missing (subject, relation, object, timestamp) quadruples.
• Includes: Link prediction (missing edges), but also entity prediction and relation prediction in a temporal context.
• Evaluation: Uses time-aware metrics like filtered temporal Hits@k and MRR.