Temporal Independent Cascade Model

• Temporal Independent Cascade Model is a diffusion framework that integrates time explicitly into the independent cascade process, capturing activation order and infection times.
• It encompasses various formulations, including discrete-time rounds, evolving-network dynamics, and hazard-based approaches, to model time-dependent influence and susceptibility.
• The model is applied to optimize interventions, infer influence via dynamic message passing, and analyze rapid spread in networks under both static and evolving conditions.

The temporal independent cascade model is a family of diffusion models that preserves the independent-cascade premise of probabilistic edgewise activation while making time an explicit modeling object. In the cited literature, temporality enters in several distinct ways: as the native discrete-round semantics of classical IC, as time-indexed transmission probabilities on evolving networks, as time-dependent hazards and influence functions, as a slowly varying global modulation of activation probabilities, and as an observation problem in which the process clock and the sampling clock need not coincide (Feng et al., 2021, Haldar et al., 2022, Gleeson et al., 2024, Chołoniewski et al., 2018, Magner et al., 2021). The unifying theme is that cascade structure is determined not only by reachability on a graph but also by infection times, temporal ordering, and the scale at which activity is observed.

1. Classical IC dynamics as a temporal process

In its classical form, the Independent Cascade model is already a temporal diffusion process on a directed graph $G=(V,E)$. A seed set $S_0\subseteq V$ is active at time $t=0$; at each subsequent discrete step, every newly active node gets one chance to activate each currently inactive out-neighbor, independently, with edge probability $p(i,j)$; once active, a node remains active, so $S_{t-1}\subseteq S_t$; and on a finite graph the process terminates in at most $n-1$ steps (Feng et al., 2021). This is the canonical one-shot, progressive, discrete-time semantics used throughout later temporal variants.

A complementary formulation represents a cascade by node activation times $\tau_i^c\in\{0,1,\dots,T\}$, with $\tau_i^c=T$ if node $i$ never activates within the observation horizon. In this representation, a node activated at time $t$ can try exactly once to activate each inactive neighbor at time $S_0\subseteq V$0, with time-homogeneous edge transmission probability $S_0\subseteq V$1 (Wilinski et al., 2020). The activation-time vector is therefore a sufficient temporal summary of a discrete IC cascade.

A more general temporal formalism replaces one-step transmission with a time-dependent hazard $S_0\subseteq V$2, interpreted as the probability that infection is transmitted along $S_0\subseteq V$3 in $S_0\subseteq V$4 after $S_0\subseteq V$5 was infected. The cumulative transmission probability is then $S_0\subseteq V$6. Discrete-time IC is recovered by the singular kernel

$S_0\subseteq V$7

which encodes a single attempt exactly one unit of time after infection (Gleeson et al., 2024). This hazard-based view places standard IC and more general temporal cascades in a common framework.

On directed acyclic graphs, the temporal IC mechanism also admits a static causal interpretation. If $S_0\subseteq V$8 denotes the active parents of node $S_0\subseteq V$9, then

$t=0$0

which is the product-form failure probability implied by independent one-shot activation attempts (Feng et al., 2021). This equivalence is central to later work on identifiability.

2. Principal temporal formulations

The literature uses the expression “temporal independent cascade model” for several non-equivalent constructions. Some retain a static network and modulate activation in time; others place IC on an evolving temporal network; others focus on time-dependent observables or interacting cascades (Chołoniewski et al., 2018, Haldar et al., 2022, Khan et al., 2023, Gleeson et al., 2024).

Formulation	Temporal ingredient	Representative use
Classical discrete-time IC	One-shot activation in rounds $t=0$1	Baseline diffusion dynamics
Hype-modulated IC	Global factor $t=0$2 modulates $t=0$3 on a static weighted network	News cascades and fluctuation scaling
Evolving-network T-IC	Time-specific probabilities $t=0$4 on $t=0$5	Disease monitoring and temporal spread optimization
Hazard-based temporal IC	Edge hazards $t=0$6 and time-dependent influence $t=0$7	Early/late influential spreaders
Interacting IC processes	Two concurrent cascades with state-dependent susceptibility	Backward and forward inference

In the evolving-network formulation, time is discrete and the network is written as $t=0$8, where each interval $t=0$9 has its own propagation map $p(i,j)$0. For a window $p(i,j)$1, the model runs a standard IC process to completion inside each interval $p(i,j)$2 using $p(i,j)$3, starting from the active set $p(i,j)$4, and then carries the resulting active set forward to $p(i,j)$5. Nodes remain active permanently, so the model effectively stacks interval-wise IC processes with time-varying edge probabilities (Haldar et al., 2022).

A different temporalization is the hype-modulated IC used for online news. There the cascade dynamics run on a largely static directed, weighted news-outlet network with base probabilities $p(i,j)$6, but each cascade receives a scalar hype parameter $p(i,j)$7 and uses

$p(i,j)$8

The crucial modeling ingredient is not merely heterogeneous $p(i,j)$9, but temporal clustering of similar hype values across consecutive cascades. In the “TP” variants, hype values are drawn from a power law and then simulated in sequence from the lowest to the highest values of $S_{t-1}\subseteq S_t$0, producing slowly varying periods of low- and high-viral potential (Chołoniewski et al., 2018).

An interacting temporal IC extends the state space itself. Each node has state

$S_{t-1}\subseteq S_t$1

and, for each directed edge $S_{t-1}\subseteq S_t$2, transmission parameters depend on whether the target has already been infected by the competing process. The parameter set for one edge is

$S_{t-1}\subseteq S_t$3

Thus infection by one process changes susceptibility to the other, and the temporal model must track two infection times $S_{t-1}\subseteq S_t$4 and $S_{t-1}\subseteq S_t$5 per node (Khan et al., 2023).

3. Temporal observables, aggregation, and clocks

Temporal IC models are defined not only by their generative rules but also by how time is observed. One line of work studies aggregate activity in observation windows of length $S_{t-1}\subseteq S_t$6. For news outlet $S_{t-1}\subseteq S_t$7, if $S_{t-1}\subseteq S_t$8 is the number of concept-related articles in time bin $S_{t-1}\subseteq S_t$9, then cumulative activity in window $n-1$0 is

$n-1$1

with mean $n-1$2 and variance $n-1$3. Temporal fluctuation scaling is then

$n-1$4

Empirically, this law holds across 11 news topics for over $n-1$5 outlets and more than $n-1$6 articles. The measured exponent $n-1$7 increases with $n-1$8, with $n-1$9 at short timescales, a crossover near $\tau_i^c\in\{0,1,\dots,T\}$0, and another at $\tau_i^c\in\{0,1,\dots,T\}$1, which the study interprets as partial synchronization for $\tau_i^c\in\{0,1,\dots,T\}$2 (Chołoniewski et al., 2018).

A second temporal issue is the distinction between process time and observation time. In the oversampled-cascade setting, the IC process evolves on intrinsic timesteps $\tau_i^c\in\{0,1,\dots,T\}$3, while the observer records a finer sequence $\tau_i^c\in\{0,1,\dots,T\}$4. A clock $\tau_i^c\in\{0,1,\dots,T\}$5 specifies how observation indices are partitioned into process steps. If $\tau_i^c\in\{0,1,\dots,T\}$6 is the true infection sequence, then $\tau_i^c\in\{0,1,\dots,T\}$7 is consistent with $\tau_i^c\in\{0,1,\dots,T\}$8 if, for every process step $\tau_i^c\in\{0,1,\dots,T\}$9,

$\tau_i^c=T$0

The distortion between two candidate clocks is defined through disagreement on pairwise temporal relations induced on vertices. Under an IC process with known parameters on Erdős–Rényi graphs, the FastClock algorithm estimates the clock in $\tau_i^c=T$1 time and attains

$\tau_i^c=T$2

with probability at least $\tau_i^c=T$3 (Magner et al., 2021).

A third observation problem is cascade reconstruction from timestamped activations. Given reported nodes

$\tau_i^c=T$4

the reconstruction task seeks a minimum-edge tree spanning all reported nodes while enforcing that every root-to-terminal path is order-respecting: a path from seed $\tau_i^c=T$5 to reported node $\tau_i^c=T$6 may not pass through a reported node $\tau_i^c=T$7 with $\tau_i^c=T$8. This yields a temporal Steiner-tree problem. The “closure” algorithm achieves a $\tau_i^c=T$9-approximation guarantee, where $i$0 is the number of active nodes, while “delayed-bfs” runs in $i$1 time and scales to very large graphs (Xiao et al., 2018). In this setting, temporal information is used as an order constraint rather than as a full probabilistic delay model.

4. Inference, learning, and identifiability

Time-resolved inference for IC often proceeds through message passing. In prediction-centric learning from partial observations, the objective is defined over activation-time marginals on observed nodes. For cascade class $i$2, observed node $i$3, and activation-time probability $i$4, the objective is

$i$5

The required temporal marginals are approximated with dynamic message passing: $i$6

$i$7

This yields a scalable learning procedure for temporal IC under partial observability and a mixture-of-models extension in which several replicas of the same graph are combined to improve temporal marginal prediction (Wilinski et al., 2020).

A different inferential target is the expected cascade size as a function of time. Under a locally tree-like approximation, the edge-level expected subtree size satisfies

$i$8

For sparse networks this gives an $i$9 algorithm for computing time-dependent influence for all seeds, distinguishing early-time from late-time influential nodes. At $t$0, the ranking reduces to degree; near criticality in the long-time limit, it converges to nonbacktracking centrality (Gleeson et al., 2024).

For interacting cascades, inference is posed from a noisy snapshot $t$1 observed at possibly unknown time $t$2. The model introduces infection times $t$3, edge activation variables, and factor-graph constraints linking them to the snapshot. Belief propagation on an inflated factor graph then produces posterior approximations for $t$4 and $t$5, enabling both backward inference on sources and forward inference on future spread (Khan et al., 2023).

Identifiability results clarify when temporal IC parameters can be recovered from observational data. In the Markovian IC model with one hidden parent $t$6 per observed node, all $t$7 are efficiently identifiable, and all incoming edge parameters $t$8 are efficiently identifiable whenever $t$9. By contrast, the semi-Markovian IC model with pairwise hidden confounders is not identifiable in general; a chain variant becomes identifiable only if enough parameters are known a priori. An IC model with a single global hidden variable is identifiable under stated structural conditions (Feng et al., 2021). These results are specific to the discrete-time one-shot temporal semantics of IC.

5. Optimization tasks and application domains

Temporal IC models are used not only to simulate cascades but also to optimize interventions and select strategically important nodes. In evolving-network T-IC, the central objective for sentinel placement is reverse spread: $S_0\subseteq V$00 where $S_0\subseteq V$01 is a realized graph over interval $S_0\subseteq V$02. The function $S_0\subseteq V$03 is submodular, so greedy maximization yields a $S_0\subseteq V$04-approximation for temporal reverse spread maximization. The same framework defines temporal expected spread maximization for identifying highly susceptible nodes, and models interventions by setting selected $S_0\subseteq V$05 to zero or reducing them over chosen intervals (Haldar et al., 2022).

The empirical scope of these optimization problems is broad. The evolving-network T-IC paper evaluates location-based networks from NYC and Tokyo, a semi-synthetic SafeGraph-trajectory dataset, the BBC Haslemere proximity data, and Italian province-level mobility data during COVID-19. In those experiments, the reverse-spread method achieves at least $S_0\subseteq V$06 higher detection success than Max-Deg in worst cases, and temporal expected spread maximization is reported as up to $S_0\subseteq V$07 better in NYC (Haldar et al., 2022). The same study frames T-IC as useful for disease monitoring, targeted restrictions, superspreader analysis, and backward contact tracing.

In news diffusion, temporal IC is used to explain aggregate synchronization rather than to identify seeds. The network is inferred from co-occurrence of publishers in cascades of similarly phrased articles, edge weights are obtained from normalized co-occurrence frequencies, and non-trivial fluctuation-scaling exponents emerge only when a heterogeneous network is combined with a slowly varying hype parameter. Without temporal hype clustering, the independent cascade model stays near the trivial exponent $S_0\subseteq V$08; with temporally clustered hype, exponents comparable to empirical long-$S_0\subseteq V$09 values appear, for example $S_0\subseteq V$10 for TP4 versus $S_0\subseteq V$11 for the unclustered P4 variant in one reported setting (Chołoniewski et al., 2018).

Time-dependent influence metrics yield a different optimization perspective. Because $S_0\subseteq V$12 is explicitly indexed by time, one may optimize for rapid spread over short horizons or for large total reach over long horizons. The reported experiments on Erdős–Rényi networks show that the identity of the best seed can change with $S_0\subseteq V$13, so “early-time influence” and “late-time influence” are not equivalent objectives (Gleeson et al., 2024). Interacting-cascade inference extends the same distinction to settings with competing or mutually modifying processes, including rumor and epidemic source detection from snapshots (Khan et al., 2023).

6. Assumptions, limitations, and conceptual boundaries

Across these works, temporal IC is not a single canonical specification but a collection of related constructions. Some models keep the graph static and make only the activation field time dependent; some evolve the contact network through $S_0\subseteq V$14; some keep the diffusion law fixed and treat temporality as an inference problem over observation clocks or snapshots (Chołoniewski et al., 2018, Haldar et al., 2022, Magner et al., 2021). A common misconception is therefore to identify “temporal IC” with any one of these formulations.

The assumptions differ accordingly. The news-fluctuation model explicitly discards timestamps as not fully reliable observables and infers a static network from content co-occurrence instead; its temporal structure is then reintroduced through cascade sequencing and windowed aggregation (Chołoniewski et al., 2018). The evolving-network T-IC preserves submodularity by stacking within-interval IC processes and keeping active nodes permanently active across intervals, which is stronger than classical one-shot activation over a single static graph (Haldar et al., 2022). The time-dependent influence metric relies on branching-process or locally tree-like assumptions and is aimed at subcritical and critical regimes rather than supercritical outbreaks (Gleeson et al., 2024). FastClock assumes known IC parameters and an oversampling regime in which every true process step has at least one observation (Magner et al., 2021). Dynamic message passing is exact on trees and asymptotically exact on sparse random graphs but becomes approximate on loopy networks (Wilinski et al., 2020).

The identifiability results likewise depend on the discrete-time IC timing assumptions: one-shot activation attempts, independence across edges, and monotone activation. The product-form conditional probabilities that make Markovian IC identifiable do not automatically carry over to repeated-attempt or non-Markovian temporal contagion (Feng et al., 2021). This suggests that “temporal IC” should be read with attention to what exactly has been temporalized: the edge set, the transmission probability, the observation schedule, the interaction between cascades, or the inferential target itself.

Taken together, the literature shows that temporal structure in IC models can serve several roles: it can govern how contagion propagates, how susceptibility changes, how activity aggregates across scales, how parameters are learned from partial data, and how source or sentinel nodes are inferred. The precise role of time is therefore the defining property of any temporal independent cascade model.