General Information Propagation Model

• General Information Propagation Model is a formal framework for describing how information, influence, and state changes propagate through structured systems.
• It encompasses diverse formulations such as temporal diffusion, hazard models, eikonal equations, and learned message passing in graph networks.
• These models enable prediction, inference, optimization, and control by linking local transmission rules with global network dynamics.

Searching arXiv for papers on general information propagation models and closely related formulations. A general information propagation model is a formal framework for describing how information, influence, activations, messages, or state changes move through a structured system. In the literature, the phrase is used for several mathematically distinct constructions: pairwise propagation probabilities on evolving social links, node-level hazard models on hidden networks, front-propagation and eikonal equations on weighted graphs, boundary-conditioned random walks with dissipation, deterministic bounded-response dynamics for influence maximization, edge-aware message passing in graph learning, and observable-based propagation bounds in interacting particle systems (Huang et al., 2013, Rodriguez et al., 2013, Dunbar et al., 2022, Stojmirović et al., 2011, Tian et al., 2021, Li et al., 2023, Schuch et al., 2010). This suggests that “general” refers less to a single canonical equation than to a common modeling ambition: specify local transmission rules, state variables, and update semantics in a way that can support prediction, inference, optimization, or control across a class of propagation phenomena.

1. Conceptual scope and principal formulations

Across the cited work, general information propagation models differ mainly in what they treat as the propagating object and how they encode causality. In social diffusion, the core quantity may be the probability that one user activates another after an exposure. In hidden-network inference, it is a node-specific hazard rate conditioned on earlier activations. In graph-based arrival-time models, it is the first time at which information reaches a node from a known boundary set. In random-walk models, it is expected visitation or absorption under dissipation and potential bias. In graph learning, it is a learned message constructed from node and edge features. In transportation and physical systems, it may be an information package or a localized observable tied to transport (Huang et al., 2013, Rodriguez et al., 2013, Stojmirović et al., 2011, Dunbar et al., 2022, Li et al., 2023, Mansour et al., 2024, Schuch et al., 2010).

Family	Representative rule	Typical setting
Pairwise temporal diffusion	$$P(\delta_{i,j,k}=1)=q_{i,j}\tau_{i,j,k}^{-\alpha_{i,j}}$$	Evolving social links (Huang et al., 2013)
Survival-theoretic diffusion	$\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$	Hidden-network inference (Rodriguez et al., 2013)
Graph arrival-time / eikonal	$\|\nabla_w^+u_i\|_p=s_i$	Front propagation on graphs (Dunbar et al., 2022)
Boundary-conditioned random walk	$$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$	Context-specific interaction networks (Stojmirović et al., 2011)
Continuous-state network dynamics	$x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$	Unified IC/LT-style influence (Tian et al., 2021)
Edge-aware graph learning	$m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$	GNN propagation (Li et al., 2023)

These formulations are not interchangeable, but they share a local-to-global structure. A local event—an exposure, activation, arrival, edge traversal, or message update—induces a system-level propagation process through repeated composition.

2. State variables, carriers, and propagation primitives

A general propagation model must specify what state is attached to nodes, edges, or links, and what event counts as successful transmission. In microscopic social diffusion, the basic label is binary: for message $k$ sent from $v_i$ to follower $v_j$,

$$\delta_{i,j,k}= \begin{cases} 1, & \text{if } v_j \text{ retweets message } k \text{ after seeing it from } v_i,\ 0, & \text{otherwise.} \end{cases}$$

The modeling target is then the pairwise propagation probability $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$0 (Huang et al., 2013).

In hidden-network diffusion models, the primitive state is the first-activation time of each node in a cascade,

$\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$1

and propagation is encoded through hazard rates whose dependence on prior activations reveals latent edges (Rodriguez et al., 2013). In deterministic continuous-state network models, each node instead carries a scalar influence level $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$2, so a node can be weakly or strongly influenced rather than merely active or inactive (Tian et al., 2021).

Other works replace adoption states by arrival times or observables. On weighted graphs, the propagated quantity is an arrival-time field $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$3, with $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$4 on known nodes and $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$5 determined by neighboring nodes with lower values (Dunbar et al., 2022). In interaction-network random walks, the central quantities are absorption probabilities and expected visit counts, computed from the transient Green’s function $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$6 under boundary conditions, potentials, and dissipation (Stojmirović et al., 2011). In interacting particle systems, propagation is tied to observables such as local particle densities

$\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$7

rather than arbitrary bounded operators on the full many-body Hilbert space (Schuch et al., 2010).

The carrier of propagation can also be explicit. In plant signaling, the input is a transmitter emission rate $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$8, the channel is a diffusion-reaction system, and the output is the number of receiver molecules $\lambda_i(t)=Y_i(t)\alpha_i(t\mid \mathbf{s}(t))$9 (Awan et al., 2018). In V2V-enabled transportation networks, carriers are vehicles classified as informed or non-informed, with counts $\|\nabla_w^+u_i\|_p=s_i$0 over spatial clusters (Kim et al., 2019). In the transportation-network IPM, the carrier is an “information package” containing a state change, a position, and a propagation rule (Mansour et al., 2024). In graph learning, the carrier is a learned message assembled from node and edge embeddings rather than a physical packet (Li et al., 2023).

3. Temporal structure, locality, and memory

One major axis of variation is how time enters the model. Some frameworks are discrete-time and iterative, some are continuous-time and event-driven, and some are formulated through equilibrium or first-arrival conditions. The temporal question is not only when propagation occurs, but what part of the past matters.

A particularly explicit temporal law appears in evolving social networks. There, the probability that $\|\nabla_w^+u_i\|_p=s_i$1 activates $\|\nabla_w^+u_i\|_p=s_i$2 depends on the latency

$\|\nabla_w^+u_i\|_p=s_i$3

defined as the idle time since the latest prior successful activation on that edge. Empirically,

$\|\nabla_w^+u_i\|_p=s_i$4

and the resulting Decay model writes

$\|\nabla_w^+u_i\|_p=s_i$5

This makes static models the special case $\|\nabla_w^+u_i\|_p=s_i$6 and treats propagation as a first-order Markov approximation in which only the time since the latest interaction affects the current decision (Huang et al., 2013).

Survival-theoretic models are also continuous-time, but they encode temporal dependence through hazards rather than explicit edge latencies. The additive model uses

$\|\nabla_w^+u_i\|_p=s_i$7

while the multiplicative model uses

$\|\nabla_w^+u_i\|_p=s_i$8

The former is purely excitatory; the latter allows both encouragement and inhibition, depending on whether $\|\nabla_w^+u_i\|_p=s_i$9 is greater or smaller than $$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$0 (Rodriguez et al., 2013).

Graph front-propagation models encode time as arrival order. In the $$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$1 formulation,

$$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$2

which is equivalent to the local eikonal equation

$$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$3

For $$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$4 and $$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$5, analogous equivalences hold between front propagation, path-set travel times, and local eikonal-type equations (Dunbar et al., 2022).

Physical and transportation models contribute a different temporal view: finite-speed or event-triggered propagation. In interacting particle systems, if particles initially occupy region $$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$6, then for any site at graph distance at least $$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$7,

$$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$8

so the signal is exponentially suppressed outside an effective light cone (Schuch et al., 2010). In V2V diffusion, the exact finite-population dynamics are a continuous-time Markov chain, while the large-scale limit is a clustered ODE system over informed and non-informed populations (Kim et al., 2019). In the IPM for transportation networks, an information package propagates until it intersects another package, reaches a boundary, or terminates, and each boundary arrival triggers a node update (Mansour et al., 2024).

4. Network structure, heterogeneity, and interaction geometry

General propagation models differ not only in dynamics but in what they assume about the underlying structure. Some assume a fixed observed graph, some infer a hidden network, and some work on multilayer, clustered, or hybrid topologies.

In hidden-network inference, the network is not observed at all; it is reconstructed from cascades by estimating edge strengths in a hazard model. The key structural assumption is progressive diffusion: nodes activate once and do not revert, while different cascades are independent (Rodriguez et al., 2013). By contrast, the continuous-state GIP model is built directly on an observed directed weighted graph $$\tilde{H}=\tilde{P}_{ST}\tilde{G},\ \tilde{F}=\tilde{G}\tilde{P}_{TS}$$9 with adjacency matrix $x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$0, and its node response function

$x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$1

unifies extended independent-cascade and extended linear-threshold dynamics while allowing feedback and repeated updates (Tian et al., 2021).

Several papers emphasize that real propagation is not tree-like. In clustered multilayer networks, one layer may represent a physical network and another an online social network, with transmissibilities $x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$2 and $x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$3 and explicit triangle structure in each layer. Increasing the level of clustering in either layer increases the epidemic threshold and decreases the final epidemic size, while low-transmissibility information spreads more effectively with a small but densely connected social network and highly transmissible information spreads better with a large but loosely connected one (Zhuang et al., 2015). Hybrid-network work makes a similar point by mixing scale-free and small-world topologies and mixing SIS, SIR, and SIRS dynamics “in arbitrary proportions,” while also introducing the blockbuster effect and implicit node edges to represent explosive spread (Nian et al., 2020).

Other frameworks make network context itself part of the model. In interaction networks, sources, sinks, potential functions, pseudosinks, and dissipation define context-specific propagation through the modified transition matrix

$x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$4

with absorbing and emitting quantities

$x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$5

This turns generic diffusion into a boundary-conditioned, potential-biased, dissipative random walk (Stojmirović et al., 2011). In graph learning, heterogeneity enters through dense and sparse feature branches and through explicit edge features. GIPA computes vector-valued attention from

$x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$6

and propagates edge-aware messages from

$x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$7

so the same source node can send different messages along different edges (Li et al., 2023).

5. Inference, prediction, optimization, and control

A central reason for making propagation models general is operational: they are used to estimate hidden parameters, predict future spread, select interventions, or design incentives. The mathematical structure of the model therefore matters as much for computation as for description.

In temporal social diffusion, the Decay model estimates $x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$8 by maximum-a-posteriori inference with priors $x_j(t)=f_{j,t}\!\left(\sum_i W_{ij}x_i(t-1)\right)$9 and $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$0. Under a “next-one” evaluation strategy, this model outperforms static baselines in retweet prediction: the best baseline AUC is $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$1, while the Decay model achieves more than $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$2, reducing the AUC-implied pairwise error from $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$3 to $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$4. In the same work, when predicted edge probabilities are fed into CELF++ for influence maximization, the selected seed set reaches $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$5 nodes versus $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$6 for the best baseline, an improvement of $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$7 (Huang et al., 2013).

Survival-theoretic models support convex likelihood-based inference. In the additive case, maximum likelihood solves a convex program over nonnegative transmission parameters; in the multiplicative case, log-reparameterization makes the problem convex in transformed variables and allows both positive and negative influence. On real cascades with more than 10 million meme cascades across 3.3 million websites, the inferred models are used not only for network reconstruction but also for forward prediction of cascade size and duration (Rodriguez et al., 2013).

Optimization enters even more directly in the unified GIP model. Influence maximization is formulated as a mixed-integer nonlinear program over initial seed values and seed indicators. In the linear special case, the exact solution is to choose the $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$8 nodes with largest $m_{i,j}^d=\tilde{\bm a}_{i,j}^d * p_{i,j}^d$9, where

$k$0

so the criterion is closely related to Katz centrality (Tian et al., 2021). Mechanism-design work pushes the same logic into incentives. PRDM defines layerwise weights

$k$1

and upward propagation rewards with factor $k$2, and is shown to be incentive compatible, asymptotically budget balanced, exactly resistant to parallel Sybil attacks, and $k$3-Sybil-proof under a majority assumption (Zheng et al., 2024).

Learning-based formulations use propagation as a trainable operator. On OGBN-proteins, full GIPA achieves test ROC-AUC $k$4, and ablations show that removing the bit-wise module, the feature-wise module, or edge features degrades performance to $k$5, $k$6, and $k$7, respectively (Li et al., 2023). This suggests that, in graph learning, “general propagation” often means relation-aware and feature-selective propagation rather than a universal diffusion law.

6. Applications, limitations, and recurring points of contention

The application range is unusually broad. Social-media retweet prediction and viral marketing are explicit targets for temporal pairwise models (Huang et al., 2013). Hidden-network survival models are applied to memes, rumors, stories, products, and diseases (Rodriguez et al., 2013). Graph arrival-time models support semi-supervised label propagation and trust ranking (Dunbar et al., 2022). Random-walk interaction models are illustrated on protein-interaction networks (Stojmirović et al., 2011). Plant communication is modeled as action-potential-modulated molecular signaling with mutual-information analysis (Awan et al., 2018). V2V-enabled transportation models treat message dissemination as a clustered epidemiological process on coupled mobility and communication graphs (Kim et al., 2019). The IPM treats transportation networks as node-link systems exchanging typed information packages (Mansour et al., 2024). In graph learning, GIPA is used for node classification on edge-attributed graphs (Li et al., 2023). In quantum and classical many-body settings, the question becomes whether physically relevant observables propagate with finite speed (Schuch et al., 2010).

Several limitations recur. Universality claims are often stronger than the empirical evidence. The temporal scaling law $k$8 is shown on Sina Weibo, and the authors present it as potentially general, but cross-platform validation is not reported in the provided text (Huang et al., 2013). Survival models are flexible and convex, but they assume progressive diffusion and require specified baseline hazards or time-shaping functions (Rodriguez et al., 2013). Compartmental SIRS-style social models are explicitly said to ignore network topology, even while acknowledging that topology affects diffusion (Sotoodeh et al., 2013). Transportation IP models are conceptually broader than traffic flow, but their explicit mathematics is traffic-specific and grounded in LWR shockwaves, node capacities, and route-choice proportions (Mansour et al., 2024). V2V models gain tractability through density-dependent limits and SI-type compartments, but they omit forgetting, recovery, and competing messages (Kim et al., 2019).

A recurring misconception is that a general information propagation model must take the form of one universal equation. The cited literature points in the opposite direction. One line of work treats propagation as a recency-weighted edge probability; another as a hazard process on hidden networks; another as a graph eikonal boundary-value problem; another as a dissipative random walk; another as continuous bounded-response dynamics; another as typed message passing in neural networks; and another as finite-speed transport of observables. This suggests that the durable common core is not a specific algebraic form, but a modeling template: define the propagating entity, define the admissible state transitions, define how locality constrains updates, and define the statistical or operational task—prediction, inference, control, optimization, or learning—that the propagation law is meant to support (Huang et al., 2013, Rodriguez et al., 2013, Dunbar et al., 2022, Stojmirović et al., 2011, Tian et al., 2021, Li et al., 2023).