HyperEvent: Higher-Order Event Modeling

• HyperEvent is a conceptual framework for modeling structured, time-stamped events that involve multiple entities instead of simple dyadic interactions.
• It uses temporal hypergraphs, directed variants, and subset repetition measures to formally represent and quantify higher-order interactions.
• Statistical models like the Relational Hyperevent Model and hyperedge-triggered frameworks enable analysis of collaborative, communicative, and bibliographic networks in dynamic settings.

HyperEvent denotes a family of concepts centered on higher-order, time-stamped events whose elementary object is not an isolated dyad but a set-valued interaction. In network science, this object appears as the relational hyperevent or hyper-event: a temporal hyperedge that may connect any number of actors, or multiple source and target sets, and whose occurrence depends on the evolving history of prior events (Lerner et al., 2 Jun 2025). In adjacent work on event-history modeling, the same idea underlies the Hyperedge Event Model, which jointly models who interacts with whom and when for one-to-many or many-to-one events (Kim et al., 2018). More recently, the term has also been used for a dynamic-link-prediction framework that treats a predicted edge as part of a cohesive composite event assembled from correlated historical interactions (Gao et al., 16 Jul 2025). Taken together, these usages suggest an umbrella notion: HyperEvent refers to modeling frameworks in which events are structured, higher-order, and temporally contextualized rather than reducible to independent pairwise ties.

1. Definition and scope

A hyper-event generalizes the dyadic relational event by allowing one event to involve more than two entities. In the temporal-hypergraph formulation, a hyperevent is a pair

$e=(t,I),$

where $t$ is the event time and $I \subseteq \mathcal{I}$ is the participating node set; a temporal hypergraph is then an ordered sequence

$E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$

This definition emphasizes that the history is itself higher-order: subsets of any size can recur, close, assort, or exhibit homophily over time (Lerner et al., 2 Jun 2025).

The same object admits directed variants. In multicast interaction models, a directed hyperevent is

$(t,i,J),$

with one sender $i$ and a receiver set $J$, while more general directed temporal hypergraphs allow events of the form

$(U_i,V_i,t_i,\delta_i),$

with disjoint source and target sets $U_i$ and $V_i$ (Lerner et al., 2021, Mellor, 2018). In bibliographic modeling, a publication may be represented as a set-to-set or tripartite hyperevent, such as

$t$0

for author and cited-work sets, or

$t$1

for authors, references, and keywords (Espinosa-Rada et al., 2024, Barbagli et al., 12 Apr 2026).

A recurring misconception is that such events can be decomposed into dyads without substantive loss. Multiple frameworks reject that premise explicitly. The Hyperedge Event Model argues that splitting multicast events into separate dyadic events distorts timing and ignores joint receiver selection (Kim et al., 2018). Relational hyperevent models make the same point for meetings, coauthor teams, and multi-actor communication, where pairwise projections mechanically create triangles and obscure genuine polyadic dependence (Lerner et al., 2019). This suggests that HyperEvent is not merely a larger event but a different ontological unit.

2. Mathematical representations

The minimal mathematical language of HyperEvent is the temporal hypergraph. For a subset $t$2, prior degree in a temporal hypergraph is defined by

$t$3

so degree is not restricted to single nodes or dyads but extends to subsets of arbitrary order (Lerner et al., 2 Jun 2025). This generalization is foundational because many HyperEvent statistics are built by summing or comparing such subset degrees.

One widely used family is subset repetition. For order $t$4,

$t$5

which yields node-level preferential attachment at $t$6, dyadic co-participation persistence at $t$7, and triadic repetition at $t$8 (Lerner et al., 2 Jun 2025). Exact group recurrence is captured separately by

$t$9

Directed and multipartite variants extend the event space rather than abandoning it. In directed hypergraphs, event-graph constructions treat events themselves as nodes in a second-order time-unfolded model, with joining rules based on set intersections such as $I \subseteq \mathcal{I}$0 for $I \subseteq \mathcal{I}$1-adjacency, or $I \subseteq \mathcal{I}$2 for walk-forming transitions (Mellor, 2018). In scientific-production models, a publication hyperedge can simultaneously encode coauthorship, citation, and keyword assignment, so that one paper induces a single tripartite event rather than a collection of separate author-author, paper-paper, and keyword-keyword dyads (Barbagli et al., 12 Apr 2026).

These representations imply that HyperEvent is fundamentally about event-space design. A plausible implication is that the main modeling choice is not only the stochastic law over time, but also which subsets, roles, and modes are treated as atomic participants of an event.

3. Statistical frameworks

The canonical statistical formulation is the Relational Hyperevent Model. In its general Cox-type form, the relative event rate for hyperedge $I \subseteq \mathcal{I}$3 at time $I \subseteq \mathcal{I}$4 is

$I \subseteq \mathcal{I}$5

with partial likelihood

$I \subseteq \mathcal{I}$6

Here $I \subseteq \mathcal{I}$7 are hyperedge statistics and $I \subseteq \mathcal{I}$8 is the risk set of candidate hyperedges (Lerner et al., 2019). The temporal-hypergraph formulation in "Modeling temporal hypergraphs" further emphasizes that, at the MLE, observed sequence-wise totals of the chosen statistics equal their expected values under the fitted model, giving RHEM a maximum-entropy or moment-matching interpretation and making it useful as a tailored null model (Lerner et al., 2 Jun 2025).

The Hyperedge Event Model is closely related but explicitly factorizes event structure and timing. It combines a dynamic ERGM-like model for receiver-set selection with a survival model for waiting times, and introduces a multivariate Bernoulli distribution with closed-form normalizing constant to ensure non-empty receiver sets (Kim et al., 2018). This formulation is particularly natural for one-to-many or many-to-one directed events, such as multicast email.

Continuous-time point-process extensions also exist. The Hyperedge-triggered Hawkes process augments a multivariate Hawkes model with a hyperedge term activated when all members of a subset co-fire within a window $I \subseteq \mathcal{I}$9: $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$0 with $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$1 and a pattern-completion anchor defined by the most recent completion time of the group (Xu, 26 May 2026). That paper derives a piecewise compensator to remove bias from naive integration, introduces a CP tensor decomposition

$E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$2

and reduces hyperedge parameterization from $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$3 to $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$4 (Xu, 26 May 2026).

Across these frameworks, the shared logic is that HyperEvent models assign event probabilities or intensities to whole hyperedges, while allowing the sufficient statistics to depend on lower-order subsets, closure structures, and event outcomes. This distinguishes them from dyadic REMs, whose event space and statistics are intrinsically pairwise.

4. Mechanisms and major application domains

The most developed applications are in collaboration, communication, and bibliographic networks. In multicast email, RHEM permits sender–receiver-set covariates such as exact repetition, unordered repetition, receiver-set heterophily, and sender-specific subset repetition, all of which are unavailable in dyadic decompositions (Lerner et al., 2021). In the Enron reanalysis, the full RHEM achieved an AIC of approximately $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$5, compared with about $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$6 for the full dyadic model, and the model recovered strong positive unordered repetition alongside negative exact repetition, consistent with turn-taking within stable participant groups (Lerner et al., 2021).

Scientific collaboration has become a major HyperEvent use case because papers are intrinsically team events. In coauthor-network models, a publication event is

$E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$7

where $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$8 is the author set and $E=[e_1=(t_1,I_1),\dots,e_M=(t_M,I_M)].$9 is a relational outcome such as normalized citations (Lerner et al., 2021). That work pairs RHEM with relational hyperevent outcome models and shows that shared prior success can increase both collaboration rates and impact, whereas some familiarity effects increase collaboration while reducing impact (Lerner et al., 2021). In the coevolutionary author–reference model, the joint analysis of 1,416,353 publication events found a strong tendency for subsets of papers to be repeatedly cited together, and the effect "cite paper and its refs" made the largest contribution among the estimated mechanisms (Lerner et al., 2023).

Group- and multipartite variants extend the same logic. The Author-Oriented Relational HyperEvent Model represents a paper as

$(t,i,J),$0

and factors the event into an author-group hazard and a conditional citation-set choice model; in Chilean astronomy, coherent groups tended to be co-cited more frequently in later publications (Espinosa-Rada et al., 2024). The tripartite model for scientific collaboration represents each paper as

$(t,i,J),$1

linking authors, references, and keywords simultaneously. In that setting the full model achieved an AIC of about $(t,i,J),$2, while removing references worsened AIC by $(t,i,J),$3, removing authors by $(t,i,J),$4, and removing keywords by $(t,i,J),$5, indicating that references carried the largest explanatory weight in that dataset (Barbagli et al., 12 Apr 2026).

Outside science, HyperEvent methods have also been used to analyze newsroom collaboration. In a dataset of 688 data-journalistic pieces with 363 distinct authors, including 29 data journalists, RHEM showed that prior common authorship affected future co-authorship and that science journalists became more prominent in post-COVID collaborative data-journalism events (Witzenberger et al., 2024). This breadth suggests that the HyperEvent formalism is portable wherever the observed unit is a team or group event rather than a dyadic act.

5. Event-centric machine learning reinterpretations

A distinct machine-learning line uses HyperEvent as a learned event-recognition framework for dynamic graphs. "HyperEvent:Learning Cohesive Events for Large-scale Dynamic Link Prediction" reframes continuous-time dynamic link prediction as hyper-event recognition: a query edge $(t,i,J),$6 is evaluated by asking whether it forms a valid hyper-event together with relevant historical events (Gao et al., 16 Jul 2025). Historical events from the endpoints’ real-time adjacency tables are transformed into an association sequence of 12-dimensional event correlation vectors built from $(t,i,J),$7-hop, $(t,i,J),$8-hop, and $(t,i,J),$9-hop neighborhood overlaps, and a Transformer discriminator predicts whether the resulting composite event is authentic (Gao et al., 16 Jul 2025). On TGB, the framework outperformed state-of-the-art methods on 4 of 5 datasets, and on the large Flight dataset it achieved a 6.95% improvement in Mean Reciprocal Rank while using only 10.17% of the training time of the state-of-the-art baseline (Gao et al., 16 Jul 2025).

An earlier event-centric bridge is the event graph, a second-order time-unfolded representation whose nodes are events and whose edges encode feasible temporal transitions. Because joining rules generalize from dyads to source and target sets, event graphs "extend easily to consider non-dyadic interactions, known as hyper-events" (Mellor, 2018). This perspective makes HyperEvent objects compatible with percolation, motif counting, centrality, and higher-order walk analysis without reducing them to dyadic paths.

Hypergraph-based event reasoning has also entered machine perception. In event-based object detection, Ev-DTAD combines Hierarchical Temporal Aggregation, a three-channel pseudo-RGB encoding of event-camera streams, with Frequency-aware Hypergraph Temporal Fusion, which uses ConvLSTM state and hypergraph message passing over multi-scale tokens (Wang et al., 9 May 2026). On Gen1, 1Mpx/Gen4, and eTraM, the model reported gains of $i$0 mAP and $i$1 faster, $i$2 mAP and $i$3 faster, and $i$4 mAP with $i$5 faster, respectively, relative to the stated baselines (Wang et al., 9 May 2026). Although this literature does not model social or bibliographic hyperevents, it shares the same design intuition: compact temporal encoding plus high-order relational aggregation.

6. Limitations, ambiguities, and open problems

The most persistent limitation is combinatorial growth. In RHEM, candidate hyperedge sets scale explosively with node count and event size, so exact denominators are generally infeasible and estimation depends on case–control sampling or analogous approximations (Lerner et al., 2019, Lerner et al., 2 Jun 2025). High-order subset statistics can also become sparse or unstable, and simultaneous or nearly simultaneous events require special handling beyond the simplest sequential formulations (Lerner et al., 2 Jun 2025).

Model misspecification remains a substantive issue. In the Hyperedge-triggered Hawkes process, pairwise parameters were recovered with relative error below 5% in synthetic experiments, but hyperedge weights showed a systematic $i$6 bias, and the bias varied non-monotonically with the kernel decay rate $i$7, ruling out a simple overlap explanation (Xu, 26 May 2026). The same study found that the inferred transition to explosive cascades occurred at about $i$8 times the theoretical critical strength, plausibly because hyperedge weights were underestimated (Xu, 26 May 2026). These results show that higher-order structure is identifiable in principle, but not automatically estimated without distortion.

Machine-learning HyperEvent formulations face different trade-offs. The dynamic-link-prediction HyperEvent model was state-of-the-art on four TGB datasets but substantially weaker on Review than GraphMixer, indicating that purely structural correlation vectors are not uniformly sufficient across domains (Gao et al., 16 Jul 2025). In event-based vision, performance depends on design choices such as window length, decay rates, inhibition strength, and the number of hyperedges, and the training procedure remains two-stage rather than end-to-end (Wang et al., 9 May 2026).

A broader conceptual limitation is terminological. The literature contains at least three non-equivalent usages: relational hyperevents on temporal hypergraphs, composite hyper-events for dynamic graph prediction, and hypergraph-based event reasoning in perception. This suggests that "HyperEvent" is presently a family resemblance term rather than a fully standardized technical designation. A plausible next step is a clearer ontology linking statistical hyperedge event-history models, second-order event representations, and learned composite-event recognizers, together with benchmarks that test whether these formalisms recover the same higher-order regularities under shared data-generating assumptions.