Temporal Interaction Networks (TINs)

• Temporal Interaction Networks (TINs) are dynamic graphs that embed interaction times and features, allowing precise analysis of causality, connectivity, and influence.
• Dynamic embedding models like JODIE use RNN-based architectures to update node trajectories and predict future interactions, showing improvements in metrics such as MRR and AUC.
• TINs facilitate applications in recommendation systems, streaming analytics, and neuroscience, while also presenting challenges in scalability and computational efficiency.

Temporal Interaction Networks (TINs) are dynamic graphs designed to represent systems where the interactions between entities are indexed by specific times and, often, by transferred quantities or associated features. This formalism has emerged as a foundational modeling tool in computational social science, neuroscience, systems biology, recommendation systems, provenance analysis, and streaming data management, among other domains. TINs generalize static networks by embedding the temporal order and potentially other attributes directly within the network structure—modifying classical notions of connectivity, causality, centrality, and influence.

1. Mathematical Formulation and Structural Properties

A Temporal Interaction Network is generally defined by a set of vertices $V$, a set of possible directed edges $E \subseteq V \times V$, and an ordered time domain $T$ (typically discrete or continuous). The fundamental data are interactions, often represented as a sequence

$I = \{(u_i, v_j, t_k, f_k)\}$

where each tuple denotes a directed interaction from node $u_i$ to node $v_j$ at time $t_k$, optionally annotated by feature vector $f_k \in \mathbb{R}^m$ (e.g., transaction amount, text content) (Kumar et al., 2019, Chen et al., 2021, Holme et al., 2011, Kosyfaki et al., 8 Jan 2026).

The topology of a TIN is inherently time-dependent: at any snapshot, the instantaneous adjacency matrix $A(t)$ encodes which edges are active, with $A_{ij}(t)$ indicating connectivity status (binary, edge weight, or flow). The network may be bipartite (as in user–item recommendation) or general directed, possibly with evolving edge weights and non-persistent links (Kumar et al., 2019, Bhaskar et al., 2023, Cao et al., 2021).

Key structural metrics and concepts include:

• Temporal degree: Number of contacts for node $i$ at time $t$.
• Reachability: Existence of time-respecting paths—a sequence of interactions $(v_0, v_1, t_1), (v_1, v_2, t_2), \ldots$ with $t_1 \le t_2 \le \ldots$ connecting $v_0$ to $v_L$ (Holme et al., 2011).
• Causality and transitivity breakdown: Unlike static graphs, the concatenation of two time-respecting paths may not lead to a valid path if intermediate timing does not align (Holme et al., 2011).

2. Dynamic Embedding, Trajectory Prediction, and Representation Learning

Dynamic representation learning in TINs aims to associate each node with a time-indexed trajectory in an embedding space, capturing both short-term and long-term evolution. Models such as JODIE (Kumar et al., 2019) instantiate coupled RNN-based architectures that update user/item dynamic embeddings at every interaction:

$$u(t) = \sigma(W_1^u u(t^-) + W_2^u i(t^-) + W_3^u f + W_4^u \Delta_u)$$

$$i(t) = \sigma(W_1^i i(t^-) + W_2^i u(t^-) + W_3^i f + W_4^i \Delta_i)$$

where $\Delta_u$ and $\Delta_i$ are time-since-last-interaction terms and $\sigma$ is an element-wise nonlinearity.

JODIE introduces a projection operator $\hat{u}(t+\tau) = [1 + w] \circ u(t)$ with $w = W_p \cdot \tau$, enabling continuous-time extrapolation of latent trajectories. This facilitates future interaction and state change prediction, with empirical results showing robust gains in MRR and AUC over baselines (Kumar et al., 2019).

Multi-relation aggregation approaches, e.g. MRATE, enhance embeddings by mining historical, common, and sequence-similarity relations, then applying hierarchical attention mechanisms (GAT and self-attention) for relation-aware propagation. The t-n-Batch algorithm provides scalable, temporally consistent mini-batch training (Chen et al., 2021).

Alternative models fuse structural graph learning and temporal point process theory, such as DSPP, which combines a bipartite-aware GNN for topological prior encoding with an attentive shift encoder for continuous-time temporal dynamics, optimized via negative log-likelihood over event intensities (Cao et al., 2021).

3. Flow, Provenance, and Pattern Mining

TINs are well-suited for modeling systems in which interactions correspond to flows of information, goods, or other quantities. The formalism is:

$G = (V, E)$

with each edge $e = (v, u)$ carrying a sequence of flows $\{(t_i, q_i)\}$ (Kosyfaki et al., 2020, Kosyfaki et al., 8 Jan 2026). Two principal flow computation models are encountered:

• Greedy flow transfer: At each interaction $(t_i, q_i)$, the source sends $\min\{q_i, B_v^{t_i}\}$ to the target; $B_v^{t_i}$ is the buffer content just before $t_i$. This can be executed in linear time, is optimal for out-degree 1 graphs, and forms the basis for practical, scalable algorithms (Kosyfaki et al., 2020).
• Maximum flow: LP or time-expanded network formulations determine the peak flow achievable through the network, respecting temporal and buffer constraints.

In provenance analytics, TINs model both discrete (identity-preserving) and liquid (aggregate/merge-able) flows, supporting query types such as backward, forward, flow lineage, and versioning provenance (Kosyfaki et al., 8 Jan 2026). State-based indexing maintains per-vertex temporal state with buffer and provenance annotations, yielding orders-of-magnitude compression compared to raw event logs and enabling efficient temporal queries.

Pattern mining in TINs involves searching for subgraph instances matching specified flow patterns, utilizing precomputations, graph simplifications, and leveraging greedy and LP-based computations for efficiency (Kosyfaki et al., 2020). Applications include financial fraud detection, transportation analytics, and streaming data provenance.

4. Generative Models and Maximum-Entropy Ensembles

Continuous-time modeling of TINs is addressed via marked point processes, with maximum-entropy temporal network (METN) ensembles factorizing the event generating process into a global time profile (e.g., Hawkes or nonhomogeneous Poisson process) and static edge probabilities:

$\lambda_{ij}(t) = \phi_r(t) w_{ij}$

with constraints on event rates and edge strengths formalized via Lagrangian optimization (Barucca, 2 Sep 2025). Log-likelihoods decompose accordingly, enabling modular parameter estimation, simulation, and integration with community/block models and motif constraints.

This framework allows for analytic computation of expected node strengths, unique edges, and provides closed-form generative algorithms for simulating event sequences. Calibration involves fitting global time processes and biproportional scaling of edge weights subject to block or structural constraints.

Extensions include integration with neural kernel parameterizations and multivariate Hawkes estimation to capture richer excitation dynamics, as well as motif-incorporated constraints for higher-order temporal structure (Barucca, 2 Sep 2025).

5. Inference, Perturbations, and Causal Structure

The inference of time-varying, directed, and possibly cyclic interaction graphs from time series data is advanced by approaches like RiTINI, which employ space-and-time attention mechanisms and graph neural ODEs to model node dynamics responsive to historical and regulatory influences (Bhaskar et al., 2023). Key architectural components:

• Space attention weights $\alpha_{ij}(t)$ and temporal attention $l_\delta$ for lagged influence modeling.
• Dynamic aggregation of node features over memory windows, yielding time-dependent edge weights capturing regulatory or functional connectivity.

Incorporating targeted node perturbations enables sharper inference of causal effects, as attention adaptively shifts to reflect system responses. Regularization terms ensure fidelity to prior graphs and sparsity, with competitive performance over static and acyclic causal inference methods.

Applications include brain functional network recovery, gene regulatory network reconstruction, and generic dynamical system modeling, with continuous-time edge inference capability exceeding classical Granger-causality or transfer entropy methods.

6. Geometric Structure, Self-Similarity, and Fractality

TINs often exhibit self-similar scaling properties under simultaneous spatial and temporal coarse-graining—termed "flow scales," parameterized by box diameter $l_b$ and time window size $\Delta t$. The resulting flow-scale fractal dimension

$$d_v = \lim_{v \to \infty} \frac{\log(N_v/N_{0,\Delta t})}{-\log(v)}, \quad v = l_b \Delta t$$

quantifies spatio-temporal self-similarity (Dutta et al., 2024). Empirical analysis reveals that some networks (e.g., Enron email interactions) manifest true flow-scale self-similarity, indicative of an underlying hyperbolic geometry with time-varying curvature $K(t) = K_0 e^{a t}$.

Simulation of TINs as point clouds evolving in dynamically curving hyperbolic space recapitulates observed scaling laws. This geometric model guides latent-space regularization and informs multiscale modeling strategies, distinguishing genuine fractal structure from mere temporal or spatial scaling. In cases lacking constant fractal dimension, embedding in variable-curvature geometries provides an explanatory mechanism for observed patterns.

7. Applications, Limitations, and Future Directions

TINs have found application in diverse domains:

• User–item recommendation: Sequential modeling of clicks, edits, bans, with continuous-time prediction and anomaly detection (Kumar et al., 2019, Cao et al., 2021, Chen et al., 2021).
• Streaming and provenance systems: Event lineage, flow tracking, and temporal state compression (Kosyfaki et al., 8 Jan 2026, Kosyfaki et al., 2020).
• Biological and neural systems: Dynamic graph inference, functional connectivity, and perturbation-driven causal analysis (Bhaskar et al., 2023).
• Social network modeling: Mechanistic simulation, motif analysis, empirical validation against face-to-face contact data (Bail et al., 2023, Holme et al., 2011).

Limitations include scalability of per-entity embeddings to massive graphs, requirement for regularization or clustering, linear projection constraints in some models, and computational cost in continuous-time ODE or attention-based architectures. Extensions to richer parametrizations, multi-party or heterogeneous interactions, motif-conditioned generative models, and distributed, adaptive indexing schemes are ongoing research directions.

A plausible implication is that as data provenance, streaming, and interaction analytics scale in complexity and volume, the principled, time-respecting structure of TINs will remain essential—not merely as a modeling convenience, but as a theoretical necessity for causal, predictive, and mechanistic understanding of dynamic systems.

---

Key references: (Kumar et al., 2019, Kosyfaki et al., 2020, Cao et al., 2021, Bhaskar et al., 2023, Dutta et al., 2024, Chen et al., 2021, Holme et al., 2011, Bail et al., 2023, Kosyfaki et al., 8 Jan 2026, Barucca, 2 Sep 2025)