Node-Driven Temporal Hypergraph Models

• Node-driven temporal hypergraph models are higher-order network formalisms where nodes and their local temporal attributes drive the formation, prediction, and generation of hyperedges.
• They employ diverse methodologies—from ego-network analysis to attributed stream models and node embedding techniques—to capture dynamic interaction patterns.
• These models demonstrate improved performance in reconstructing temporal orders and forecasting hyperedge events while accounting for state persistence, memory, and heterogeneity.

Node-driven temporal hypergraph models are temporal higher-order network formalisms in which the primary source of structure is attached to nodes: a node’s local interaction history, latent activity state, attribute trajectory, embedding neighborhood, or event propensity determines how time-varying hyperedges are represented, predicted, or generated. In this literature, the node-centric perspective appears in several distinct forms: local temporal views such as hypergraph ego-networks, attributed stream models with temporal stars, predictive architectures that construct hyperedges from node representations, and generative models in which system-level hyperedge activity emerges from independently evolving nodes (Comrie et al., 2021, Failla et al., 2023, Mancastroppa et al., 1 Jul 2025, Jo et al., 9 Apr 2026).

1. Conceptual scope and model families

A common feature of these models is that higher-order temporal structure is not treated as a fixed sequence of exogenous group events only. Instead, the temporal hypergraph is organized around nodes and their roles in the dynamics. In some cases, the node-centered object is explicitly local, as in ego-networks and temporal stars. In others, hyperedges are learned from node embeddings, induced from node neighborhoods, or generated from node-state dynamics. This suggests a broad but coherent class of models in which node-level information is the operative unit for representing or explaining temporal higher-order interactions (Comrie et al., 2021, Zhao et al., 2023, Liu et al., 18 Jun 2025).

Family	Node-driven mechanism	Representative paper
Local temporal representation	Sequence of simplices involving an ego or its alters	(Comrie et al., 2021)
Attributed stream formalism	Temporal nodes, temporal hyperedges, and time-varying node attributes	(Failla et al., 2023)
Higher-order pattern prediction	Historical context extracted around a node triplet of interest	(Liu et al., 2021)
Dynamic graph augmentation	Hyperedges built from node embeddings across time slices	(Ma et al., 2024)
Structure learning	Incidence matrix learned from node states	(Zhao et al., 2023)
Directed event forecasting	Node event forecasting reduces hyperedge search space	(Gracious et al., 2023)
Heterogeneous temporal hypergraphs	Hyperedges induced from evolving heterogeneous node neighborhoods	(Liu et al., 18 Jun 2025)
Node-first generators	Hyperedge activity emerges from node activity and memory	(Mancastroppa et al., 1 Jul 2025, Jo et al., 9 Apr 2026)

The scope of the topic therefore spans representation, prediction, and generation. Some models begin from temporal hypergraphs as the native data structure; others begin from dynamic graphs and construct temporal hypergraphs as latent or auxiliary structures. A plausible implication is that the term node-driven names a modeling principle rather than a single architecture: nodes are the locus at which temporal dependence, semantic heterogeneity, and higher-order interaction structure are made computationally accessible.

2. Node-centered local temporal objects

The most explicit node-centered formulation is the hypergraph ego-network introduced in "Hypergraph Ego-networks and Their Temporal Evolution" (Comrie et al., 2021). In a temporal hypergraph with timestamped simplices $S$, the local evolution around an ego $u$ is represented by higher-order interactions involving $u$ and its alters. Three nested variants are defined: $T(u)=\{\, s \in S : u\in s \,\},$

$R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$

$C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$

with

$T(u)\subseteq R(u)\subseteq C(u).$

The benchmark problem is temporal reconstruction: given an ego-network with $m$ hyperedges but with the order removed, recover the original temporal ordering. Reconstruction quality is measured by the fraction of pairwise precedence relations that are correct; random guessing gives about $50\%$. The paper reports recurrent local signals used for reconstruction, including large overlap between consecutive simplices, temporally dense alter-networks, predictive ego-arrival time in radial and contracted networks, regular novelty rates, and earlier appearance of high-degree alters.

A related but more general node-centric object appears in Attributed Stream Hypergraphs (ASHs), defined as

$\mathcal{S}=(T,V,W,E,L),$

where $u$0 is the set of discrete time instants, $u$1 the set of nodes, $u$2 the temporal nodes, $u$3 the temporal hyperedges, and $u$4 the set of node attributes (Failla et al., 2023). The temporal star of node $u$5 at time $u$6,

$u$7

is the basis for node-centric temporal analysis. ASH further defines star homogeneity as a node-level homophily measure, distinct from hyperedge purity: purity asks how homogeneous the group is, whereas star homogeneity asks whether a node tends to participate in groups whose dominant label matches its own label. The distinction is central to the paper’s analysis, and it formalizes a point often obscured in pairwise temporal networks: a locally homogeneous group structure does not imply homogeneous node embedding in time.

These formulations anchor node-driven temporal hypergraph modeling in local observables. Rather than beginning with the entire evolving hypergraph, they isolate a node-centered temporal environment and study how that environment grows, stabilizes, or changes.

3. Predictive models based on node histories and node-induced hyperedges

Several neural models operationalize the node-driven view by predicting higher-order temporal structure from node-level context. HIT, introduced in "Neural Predicting Higher-order Patterns in Temporal Networks," focuses on a node triplet of interest $u$8 and predicts whether the future realizes as Edge, Wedge, Triangle, or Closure (Liu et al., 2021). Historical structure is captured by temporal random walks sampled backward from $u$9, $u$0, and $u$1; node identities are replaced by an asymmetric distance encoding that keeps $u$2 and $u$3 symmetric with respect to each other but not with respect to $u$4. The encoded walks are processed by an RNN with learnable Fourier time features and pooled to node embeddings $u$5. The model addresses three tasks: predicting interaction type, predicting when a given pattern first appears, and ranking discriminatory temporal random walks for interpretability. The paper reports average AUC gains of about $u$6 over heuristic and neural baselines for identifying interaction type, as well as uniformly better time estimation.

HYDG transfers the same principle to dynamic graph node classification by building temporal hypergraphs over node embeddings rather than over explicitly observed higher-order events (Ma et al., 2024). For each snapshot $u$7, a GNN backbone produces $u$8. The individual-level hypergraph connects a node $u$9 to the $T(u)=\{\, s \in S : u\in s \,\},$0 nearest nodes from other time slices within $T(u)=\{\, s \in S : u\in s \,\},$1, with separate short-term, mid-term, and long-term threshold ranges. The group-level hypergraph clusters node embeddings by true class label and time slice, aggregates each cluster by $T(u)=\{\, s \in S : u\in s \,\},$2, $T(u)=\{\, s \in S : u\in s \,\},$3, or $T(u)=\{\, s \in S : u\in s \,\},$4, and then applies the same hyperedge-building mechanism to the resulting class-group representations. Hypergraph propagation uses weighted node-to-hyperedge aggregation and hyperedge-to-node reweighting, and training combines individual and group losses through

$T(u)=\{\, s \in S : u\in s \,\},$5

The paper reports that HYDG consistently outperforms baselines across five dynamic graph datasets, with a typical improvement of about $T(u)=\{\, s \in S : u\in s \,\},$6–$T(u)=\{\, s \in S : u\in s \,\},$7 accuracy on DBLP5 and Reddit over the best baselines.

DyHSL goes further by learning the temporal hypergraph incidence structure directly from node states (Zhao et al., 2023). In a temporal graph whose nodes are time-location observations $T(u)=\{\, s \in S : u\in s \,\},$8, the incidence matrix is learned as

$T(u)=\{\, s \in S : u\in s \,\},$9

where $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$0 stacks node state representations and $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$1 is learnable. Hyperedge embeddings are computed from incident node states, and node embeddings are then reconstructed from associated hyperedges through node $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$2 hyperedge $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$3 node message passing. DyHSL complements this with Interactive Graph Convolution for neighbor-pair interactions and a multi-scale temporal module using $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$4. On PEMS03, PEMS04, PEMS07, and PEMS08, the paper reports best performance on all datasets and all metrics, with fewer parameters than DSTAGNN and comparable runtime.

Across these models, the predictive pattern is consistent: node-centered temporal context is first encoded, and higher-order temporal predictions are made from that context rather than from static hyperedge templates.

4. Generative and mechanistic node-first models

A different branch of node-driven temporal hypergraph research is generative. EATH, the Emerging Activity Temporal Hypergraph model, starts from node-level latent activity dynamics rather than from hyperedges as primitive objects (Mancastroppa et al., 1 Jul 2025). Each node $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$5 has persistence activity $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$6, instantaneous activity $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$7, and order propensity $R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$8, with

$R(u)=\{\, s\in S : s \subseteq A(u)\cup\{u\}\,\},$9

Nodes independently alternate between low-activity and high-activity phases, with instantaneous activity

$C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$0

where $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$1. System activity emerges from node activity through

$C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$2

Hyperedges are then generated by combining activity, order propensity, long-term memory through a frozen pairwise matrix $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$3, and short-term memory through hyperedge persistence or gain/loss of one node. The model is calibrated from empirical parameters such as $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$4, $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$5, $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$6, $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$7, $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$8, $C(u)=\{\, s \cap (A(u)\cup\{u\}) : s\in S \,\},$9, $T(u)\subseteq R(u)\subseteq C(u).$0, $T(u)\subseteq R(u)\subseteq C(u).$1, and long-term memory heterogeneity. The paper shows that EATH reproduces global activity timelines, hyperedge size distributions, node and hyperedge temporal statistics, projected graph structure, and higher-order contagion outcomes, while the memoryless baseline EATHw fails to reproduce hyperedge-level temporal structure realistically.

"Modeling non-Poissonian temporal hypergraphs by Markovian node dynamics" provides a more analytically tractable mechanism (Jo et al., 9 Apr 2026). Each node has a binary state

$T(u)\subseteq R(u)\subseteq C(u).$2

evolving independently via a two-state Markov chain with transition probabilities $T(u)\subseteq R(u)\subseteq C(u).$3 and $T(u)\subseteq R(u)\subseteq C(u).$4. A hyperedge event probability depends on the number $T(u)\subseteq R(u)\subseteq C(u).$5 of incident high-state nodes, with two rules: the AND rule, under which the hyperedge is high only if all nodes are in state $T(u)\subseteq R(u)\subseteq C(u).$6, and the LIN rule, under which the event probability is linear in the fraction of high-state nodes. Despite Markovian node dynamics, the model analytically derives interevent time distributions and autocorrelation functions that are mixtures of geometric components. The central conclusion is that Markovian latent node dynamics can generate longer-tailed interevent time distributions and slowly decaying autocorrelation in observed hyperedge and node event sequences. The paper further shows that average event probability and the coefficient of variation typically decrease with hyperedge size $T(u)\subseteq R(u)\subseteq C(u).$7, a pattern that largely agrees with several empirical datasets and is more consistent with the AND rule than with the LIN rule.

These generative models share a strong mechanistic claim: realistic temporal hypergraph phenomena can emerge from node-level heterogeneity, state persistence, and memory, without specifying the full hyperedge process as a primitive object.

5. Attributes, directionality, and heterogeneous node roles

Node-driven temporal hypergraph models also extend to settings in which node attributes, edge directionality, and heterogeneous types are structurally essential. In ASHs, attributes are explicit and time-indexed through $T(u)\subseteq R(u)\subseteq C(u).$8, but the formalism treats them as part of the representation rather than as causal generators of temporal dynamics (Failla et al., 2023). The temporal evolution is encoded by $T(u)\subseteq R(u)\subseteq C(u).$9, $m$0, and $m$1; attributes are observed alongside this evolution and used for analysis through consistency, purity, entropy, and star homogeneity. This directly corrects a common misconception: in the ASH formal model, node attributes do not drive edge formation.

DHyperNodeTPP addresses an orthogonal extension: directed higher-order temporal relations (Gracious et al., 2023). A directed hyperedge is

$m$2

with a right hyperedge $m$3 and a left hyperedge $m$4. The central difficulty is that the number of possible hyperedges is exponential in the number of nodes. The model addresses this through a three-stage sequential generative framework: node event forecasting with a temporal point process predicts which source-side nodes become active and when; candidate generation predicts right and left hyperedge sizes together with projected adjacency vectors; directed hyperedge prediction then searches only within the reduced candidate set. This is explicitly node-driven because hyperedge forecasting is conditioned on node event forecasting. The temporal-node stage models inter-event times as Lognormal, node memory is updated with a GRU, recent right and left higher-order neighborhoods are encoded with multi-head attention, and the final hyperedge scorer uses cross-attention and self-attention across the two sides of the directed hyperedge. The paper reports performance gains over state-of-the-art pairwise and hyperedge event forecasting models for event type prediction.

HTHGN generalizes the node-driven idea to heterogeneous temporal graphs by constructing heterogeneous temporal hypergraphs from node neighborhoods rather than relying on expert annotations or meta-paths (Liu et al., 18 Jun 2025). Starting from snapshots $m$5, it defines $m$6-hop and $m$7-ring hyperedges around nodes and then constrains them to $m$8-uniform size. Hyperedges become first-class nodes through heterogeneous star expansion, enabling temporal message passing between heterogeneous nodes and hyperedges. Type-specific feature projection, relationship-aware multi-head attention, semantic attention across relation types, temporal attention across snapshots, and a contrastive objective for low-order structural ambiguity jointly yield final node embeddings used for dynamic link prediction and new link prediction. The paper reports best scores on Yelp, DBLP, and AMiner for both tasks.

Taken together, these extensions show that node-driven temporal hypergraph modeling is compatible with several forms of semantic enrichment. Directionality, heterogeneity, and attributes do not remove the node-centric principle; they specify how node roles are differentiated.

6. Benchmarks, empirical regularities, and limitations

The empirical literature has converged on several benchmark styles. One is reconstruction of local temporal order from unordered higher-order interactions, as in ego-network temporal reconstruction, where performance is measured by the fraction of correctly ordered pairs and random orderings yield about $m$9 (Comrie et al., 2021). Another is predictive classification of higher-order outcomes, as in HIT’s Edge/Wedge/Triangle/Closure task. Others include node classification on future snapshots, traffic forecasting, dynamic link prediction, new link prediction, time-respecting path analysis, and higher-order contagion simulation (Liu et al., 2021, Ma et al., 2024, Zhao et al., 2023, Liu et al., 18 Jun 2025, Mancastroppa et al., 1 Jul 2025).

Several regularities recur across domains. In ego-networks, adjacent simplices often overlap heavily, alter activity is temporally localized, ego-arrival time is predictive in radial and contracted networks, and high-degree alters tend to appear earlier (Comrie et al., 2021). In ASHs, group homophily and node-centric embedding need not coincide: hyperedge purity can be high while star homogeneity remains lower, implying that individual users may participate in a heterogeneous mix of groups even when group-level polarization is strong (Failla et al., 2023). In node-state generative models, longer-tailed interevent time distributions and slow autocorrelation decay can arise from mixtures over latent node states rather than from intrinsically non-Markovian event rules (Jo et al., 9 Apr 2026).

The limitations are equally instructive. HTHGN assumes a snapshot-based temporal setting and induces hyperedges from topology, so the quality of the constructed hyperedges depends on neighborhood structure; the authors also state that interpretability of its effectiveness remains open (Liu et al., 18 Jun 2025). The Markovian node-dynamics model notes that inference is difficult, periodic structure is absent, only two node states are considered, and the hyperedge event mechanism is idealized (Jo et al., 9 Apr 2026). HYDG uses the group-level hypergraph only during training because labels are unavailable at test time, which constrains how class-level information can be operationalized in deployment (Ma et al., 2024). DyHSL’s results and ablations show that dynamic structure learning, interactive graph convolution, and multi-scale temporal modeling are all necessary; removing any of them increases prediction error (Zhao et al., 2023).

A final misconception clarified by these works is that node-driven modeling is not equivalent to same-node sequence modeling. HYDG is explicit on this point: RNNs and self-attention that only aggregate a node’s own history are too restrictive for graphs where nodes change features and labels, disappear, or newly appear (Ma et al., 2024). In the broader temporal hypergraph literature, node-driven models instead let a node aggregate or generate information through multiple related nodes, multiple time ranges, and higher-order structures. This suggests that the distinctive contribution of the paradigm is not simply temporal recurrence at the node level, but node-centered access to evolving higher-order context.