Independent Cascade Model

• Independent Cascade is a probabilistic model describing influence spread in networks where activated nodes attempt to activate each neighbor exactly once.
• It distinguishes between edge-level and node-level feedback, facilitating precise parameter estimation and the design of regret-optimal online learning algorithms.
• Recent advances extend the IC framework to incorporate realistic feedback regimes, causal inference, and efficient exploration-exploitation strategies.

The Independent Cascade (IC) model is a seminal framework for modeling influence propagation and cascade dynamics in networks. It is extensively utilized across social network analysis, viral marketing, information diffusion, and epidemiological modeling due to its operational simplicity, rigorous probabilistic semantics, and suitability for optimization. Over time, the foundational IC process has been extended to address realistic feedback regimes, causal identifiability, and practical online learning. The following sections provide a detailed exposition of the formal IC process, the distinction between edge- and node-level feedback, the statistical learning procedures for model identification, the design and analysis of confidence sets and regret-optimal online learning algorithms, and the wider implications for causal inference and influence maximization.

1. Formal Definition of the Independent Cascade Diffusion Process

The IC model operates on a directed (or undirected) graph $G = (V, E)$, where each edge $e = (u, v)$ is parameterized by an unknown activation (or “influence”) probability $p_{u,v} \in (0,1)$. Given a seed set $S \subseteq V$ activated at $t = 0$, diffusion unfolds in discrete rounds $t = 1, 2, ...$:

• Each node $u$ activated at time $t-1$ receives a single, independent attempt to activate each currently inactive out-neighbor $v$, succeeding with probability $p_{u,v}$.
• An activated node enters and remains in the active state permanently; no edge is used for propagation more than once.
• The process continues until no further activations occur.

The IC process can equivalently be described through a “live-edge” graph representation: pre-sample an independent Bernoulli $e = (u, v)$0 on each edge. The subgraph formed from the edges where $e = (u, v)$1 ($e = (u, v)$2) contains all possible influence channels, and the final set of active nodes is precisely those reachable from $e = (u, v)$3 in $e = (u, v)$4 (Yang et al., 2021).

2. Feedback Regimes: Edge-Level vs. Node-Level

Proper learning and optimization in the IC framework require feedback regarding which nodes are activated and (ideally) which edges succeeded in transmitting influence. Two primary regimes are defined:

• Edge-level feedback: After each round, the learner observes, for every active parent $e = (u, v)$5 of a node $e = (u, v)$6, whether the edge $e = (u, v)$7 succeeded (transmitted activation) or failed. This enables direct estimation of each $e = (u, v)$8.
• Node-level feedback: The learner observes only which nodes became newly active per round, with no information on which parent (from potentially many) caused each activation. In this case, if $e = (u, v)$9 is the set of active parents at time $p_{u,v} \in (0,1)$0, the activation probability is aggregated as

$p_{u,v} \in (0,1)$1

The observed activation $p_{u,v} \in (0,1)$2 satisfies $p_{u,v} \in (0,1)$3 with probability $p_{u,v} \in (0,1)$4, but the individual causal edge remains unobserved (Yang et al., 2021).

Node-level feedback is considerably more challenging, due to both nonconvex likelihood aggregation and the inherent “censoring” of the true activation pathway.

3. Parameter Estimation: Likelihood, Re-Parameterization, and Confidence Sets

Parameter recovery is fundamentally an estimation problem. The typical approach is maximum likelihood estimation (MLE) over observed activations. Due to the nonlinear aggregation, this is facilitated by a reparameterization:

• Let $p_{u,v} \in (0,1)$5, so $p_{u,v} \in (0,1)$6.
• Suppose further a linear parametrization $p_{u,v} \in (0,1)$7 for features $p_{u,v} \in (0,1)$8.

For node-level feedback, each observed attempt at $p_{u,v} \in (0,1)$9 yields a sample $S \subseteq V$0, where $S \subseteq V$1, with $S \subseteq V$2. The collection of observed data forms the log-likelihood: $S \subseteq V$3 This likelihood is concave in $S \subseteq V$4 due to $S \subseteq V$5, ensuring a unique maximizer $S \subseteq V$6.

For valid statistical inference, confidence ellipsoids are constructed: $S \subseteq V$7 where $S \subseteq V$8 and $S \subseteq V$9 depends on problem-dependent constants and the sample size. The main result (Theorem 1 of (Yang et al., 2021)) provides nonasymptotic concentration of $t = 0$0 around the true parameter $t = 0$1 under mild regularity assumptions.

4. Two-Phase Online Learning: Exploration, Exploitation, and Regret Analysis

To efficiently learn IC parameters and maximize influence in an online setting, a two-phase algorithm (TPNodeIM) is employed:

• Exploration phase: $t = 0$2 rounds where a set of “basis” edges with linearly independent features are sequentially probed, injecting enough information for parameter identifiability by ensuring $t = 0$3 and $t = 0$4 are well-conditioned in all directions.
• Exploitation phase: Over the remaining rounds, the current parameter confidence set is used to select approximately optimal seed sets via an offline $t = 0$5-oracle that returns a seed set $t = 0$6 and a parameter $t = 0$7, guaranteeing, with probability at least $t = 0$8, that the expected spread under $t = 0$9 is at least an $t = 1, 2, ...$0 fraction of the optimal over the current confidence region.

Regret is quantified in “$t = 1, 2, ...$1-scaled” terms: $t = 1, 2, ...$2 where $t = 1, 2, ...$3 is the expected spread of $t = 1, 2, ...$4 under $t = 1, 2, ...$5. The cumulative regret is shown to be $t = 1, 2, ...$6 up to problem-dependent constants, matching the optimal minimax regret for the edge-level feedback scenario (with an extra logarithmic factor for exploration) (Yang et al., 2021).

Key proof mechanisms include:

• Matrix concentration to ensure that the empirical feature matrices $t = 1, 2, ...$7 cover the relevant parameter space,
• High-probability coverage of $t = 1, 2, ...$8 by the confidence ellipsoids,
• Lipschitz-type control of the regret in terms of parameter error and path-relevant features over possible cascades.

5. Comparison with Edge-Level Feedback and Implications

Results demonstrate that, despite the weaker information content in node-level feedback (i.e., censored, nonlinear, aggregated observations), both the parameter estimation and online learning performance can match the minimax rates obtainable under edge-level feedback, modulo minor log-factors (Yang et al., 2021). In the edge-level case, simple UCB or Thompson Sampling methods directly estimate each $t = 1, 2, ...$9 (or its feature-parametric analogue) and achieve $u$0 regret without needing a distinct exploration phase. Node-level feedback necessitates a careful design of exploratory actions to address the identifiability and curvature challenges imposed by censored, nonlinear observations.

This analysis closes a major theoretical gap regarding online learning under realistic observation regimes in IC-based influence maximization.

6. Identifiability, Causal Extensions, and Applications

The issue of identifiability—determining whether model parameters can be uniquely recovered from observational data—has significant implications for influence maximization and causal inference. Extended versions of IC, incorporating unobserved confounders, are analyzed as causal graphical models. Important cases include:

• Markovian IC: Unique hidden parent for each observed node; parameters are identifiable.
• Semi-Markovian IC: Hidden pairwise confounders; generally unidentifiable without additional constraints or priors.
• Global hidden variable IC: Single global confounder; identifiable under conditions on disconnected triplets (Feng et al., 2021).

If identifiability fails, any learned parameter set is observationally equivalent to multiple true processes, rendering naive influence maximization unreliable. When identifiability holds, one can consistently estimate propagation probabilities and deploy classical optimization and bandit algorithms with theoretical guarantees.

7. Outlook and Open Problems

Current online IC learning under node-level feedback establishes both confidence-region construction and regret matching performance. Notably, key unresolved directions concern:

• Nonparametric estimation without exponential or linear-form parameterizations;
• Handling time-varying, nonstationary, or adversarially evolving networks;
• Designing seed-selection strategies under dynamic or nonstationary feedback.

These challenges underscore the importance of integrating statistical learning theory, optimization, and causal inference in further developing the IC modeling framework for realistic applications in networked systems (Yang et al., 2021, Feng et al., 2021).