Temporal Event Graphs: Causal Network Analysis

Updated 9 May 2026

Temporal Event Graphs (TEGs) are static, directed structures that uniquely encode the complete temporal and causal relationships of timestamped events.
They facilitate efficient analysis of higher-order patterns, including motif counting, percolation transitions, and causal reasoning using standard graph algorithms.
TEGs offer a lossless and scalable representation, enabling extensions to spatio-temporal, hierarchical, and process-driven networks for advanced dynamic analysis.

A Temporal Event Graph (TEG) is a static, directed, and often acyclic representation encoding the full causal and temporal structure of a dynamic network whose edges are timestamped events, such as email interactions, communications, or physical contacts. In a TEG, nodes correspond to individual events (e.g., $e_i=(u_i,v_i,t_i)$ ), and directed edges encode possible "temporal adjacency" relationships between events—specifically, those necessary to preserve all time-respecting paths and vital temporal motifs of the underlying network. This construction allows exact encoding of the temporal, causal, and often semantic structure of the original event sequence, enabling efficient analysis of higher-order temporal patterns, path-based spreading, causal reasoning, and more, using standard (static) graph-theoretic algorithms (Mellor, 2017, Saramäki et al., 2019, Mellor, 2018, Heeg et al., 30 May 2025).

1. Formal Definitions and Construction

A Temporal Event Graph is built from a temporal network $G = (V, E, T)$ where $V$ is the set of entities, $T$ the time domain, and $E$ a set of timestamped edges (events), typically of the form $e_i = (u_i, v_i, t_i)$ . Each such event becomes a node in the TEG. Directed edges between event nodes capture temporal adjacency—commonly defined as follows:

Two events $e_i$ $e_{i}$ and $e_j$ $e_{j}$ are connected by a directed edge $e_i \rightarrow e_j$ $e_{i} \to e_{j}$ if:
- $\{u_i, v_i\} \cap \{u_j, v_j\} \neq \emptyset$ (they share at least one node)
- $G = (V, E, T)$ 0 for a chosen temporal adjaceny parameter $G = (V, E, T)$ 1

All such edges are weighted by the inter-event time $G = (V, E, T)$ 2. The resulting graph $G = (V, E, T)$ 3 is a static directed acyclic graph (DAG) encoding all time-respecting possibilities in the original temporal network (Mellor, 2017, Kivelä et al., 2017, Saramäki et al., 2019, Mellor, 2018).

For directed or process-driven networks, stricter adjacency rules (e.g., the target of $G = (V, E, T)$ 4 is the source of $G = (V, E, T)$ 5) are imposed, and for events with duration or non-dyadic/hyperedge cases, the formalism extends by adapting adjacency and labeling rules (Mellor, 2018).

2. Theoretical Properties: Static, Lossless, and Unique Representation

The TEG offers three unique representational guarantees:

Static: The full temporal structure is mapped to a static graph, avoiding the need for time-sliced layers.
Lossless and Unique: The labeled TEG—a DAG where each edge carries its inter-event-time (IET) and motif—uniquely and bijectively encodes the original temporal network (up to node and time relabeling). Conditions including consistent edge sums and motif switching guarantee no two distinct temporal networks share the same TEG (Mellor, 2017).
Canonical Decomposition: TEGs admit a decomposition into temporal components (weakly connected components under $G = (V, E, T)$ 6-adjacency) that encode episodes, cascades, or collective bursts at scale (Mellor, 2018, Mellor, 2017).

The TEG thus functions as a lossless memory structure preserving all time-respecting walks and higher-order path information, unlike time-aggregated graphs or discrete snapshots which collapse or obscure critical temporal order (Mellor, 2017, Heeg et al., 30 May 2025).

3. Algorithmic Construction, Complexity, and Variants

Efficient construction of TEGs operates via:

Sorting events by time: $G = (V, E, T)$ 7 for $G = (V, E, T)$ 8 events.
For each event, finding all future $G = (V, E, T)$ 9-adjacent events (per node): $V$ 0.
Edge-pruning for thresholded TEGs is trivial ( $V$ 1 for each threshold).
Motif and IET labeling, component decomposition, and per-component analysis are $V$ 2 once the adjacency DAG is built (Mellor, 2017, Kivelä et al., 2017, Mellor, 2018, Saramäki et al., 2019).

Several variants exist:

Weighted TEGs: Edge weights explicitly encode inter-event times, directly supporting shortest-path and percolation analyses (Kivelä et al., 2017, Saramäki et al., 2019).
Directed walk-forming TEGs: Enforce stricter causal adjacency for process flows or semi-Markov dynamics (Mellor, 2018).
P-time Event Graphs: Event graphs where transitions are subject to place-specific sojourn time intervals, analyzable via max-plus algebra (Zorzenon et al., 2022, Zorzenon et al., 2021).
Hierarchical TEGs: Multilevel event graphs constructed via self-supervised pooling, as in TimeGraphs, for multi-scale event abstraction and efficient streaming inference (Maheshwari et al., 2024).

4. Key Applications: Temporal Percolation, Motif Analysis, and Causal Reasoning

Applications of TEGs are numerous in temporal network science:

Temporal-network percolation: TEGs naturally encode all finite $V$ 3-constrained time-respecting paths. Varying $V$ 4 yields percolation transitions in component sizes, interpreted analogously to directed percolation criticality; critical waiting times can be empirically determined for spreading processes in real-world networks (e.g., SI models on calls, flights) (Kivelä et al., 2017, Saramäki et al., 2019).
Temporal motifs: TEGs support efficient motif counting—not just of static subgraphs but of order- $V$ 5 local event sequences, distinguishing, for example, reciprocation, broadcasting, and sequential transfer motifs (ABBA, ABCA, etc.) and their associated inter-event time distributions (Mellor, 2017, Mellor, 2018).
Collective behavior and component-level analytics: Weakly connected components in TEGs correspond to temporally localized bursts (conversations, cascades). Embeddings based on motif-vector, IET-entropy, and topological descriptors enable unsupervised clustering and behavioral taxonomy (Mellor, 2018).
Causal structure and generalization: The TEG’s static, path-centric topology exactly preserves temporal causality. Modern results establish that isomorphism of TEGs precisely corresponds to isomorphism of time-respecting path structure. This is leveraged to generalize WL-isomorphism tests and to define expressive neural message passing schemes provably sensitive to causal topologies (Heeg et al., 30 May 2025).
Event temporal relation extraction: In NLP, TEGs with labeled temporal relations (before, after, includes, is_included, simultaneous) model rich event-event structure. Graph generation tasks employ auto-regressive LLMs with set-aligned training to avoid order-penalty artifacts, explicitly modeling the TEG as a set-valued structure (Tan et al., 2024).

5. Extensions: From Knowledge Graphs to Hierarchical and Spatio-Temporal Event Graphs

Recent work generalizes TEGs to other settings:

Temporal Event Knowledge Graphs (tEKGs): For process mining and business analytics, tEKGs extend Event Knowledge Graphs by introducing time-indexed snapshot nodes for each object, every time any attribute changes. This supports queries over temporally-localized object states and event sequences—enabling analysis unfeasible in static models (Khayatbashi et al., 2024).
Spatio-temporal event graphs: Integration of semantic, spatial, and temporal context in event graphs supports robust causal discovery from, e.g., disaster-related social media data. Nodes are enriched with semantic (LLM-based), temporal, and spatial features; edge formation is context-aware; and analysis is performed using context-sensitive GATs, yielding improved recall and interpretability (Duong et al., 1 Feb 2026).
Hierarchical time-adaptive TEGs: Construction of multilevel TEGs via data-driven pooling and cross-level message passing provides efficiency and adaptivity (jumping to coarser event abstractions during low-activity periods), with applications in dynamic event prediction and recognition (Maheshwari et al., 2024).

6. Expressiveness, Limitations, and Open Problems

TEGs are maximally expressive with respect to temporal causality: all time-respecting paths (within the adjacency rule) are retained, and the structure is static and lossless (Mellor, 2017, Heeg et al., 30 May 2025). However, certain limitations and open questions remain:

Parameter selection: The temporal adjacency window $V$ 6 fundamentally affects component structure; heuristic or domain-specific approaches are often used.
Motif analysis: Standard TEGs focus on two-event motifs; rigorous null-models or higher-order (e.g., $V$ 7) extensions are open topics (Mellor, 2017, Mellor, 2018).
Handling of non-dyadic interactions: Complete generalization to hyper-events and higher-order dynamical motifs is possible but requires careful adjacency specification (Mellor, 2018).
Scalability in dense or attribute-rich settings: Pairwise expansion for temporal snapshots (in tEKG) or high-frequency event streams can render naive algorithms quadratic, though practical optimizations tend to keep performance near-linear (Khayatbashi et al., 2024, Maheshwari et al., 2024).
Expressivity-vs-cost in neural architectures: Message-passing on augmented TEGs is theoretically maximally expressive for causal topology but may be computationally intensive for extremely large temporal event sets (Heeg et al., 30 May 2025).

TEGs are thus central to contemporary research in temporal networks, process mining, information diffusion, temporal reasoning, and explainable event-driven AI. They form a robust backbone for both theoretical analysis and practical, scalable computation in settings where time and causality are fundamental.