Source-linked AI summary
From Content Sharing to Opinion Shaping: A Theoretical Framework for Cascade-Driven Opinion Dynamics
Elisabetta Biondi, Chiara Boldrini, Andrea Passarella, Marco Conti
TL;DR
Opinion formation and information diffusion co-occur in online social networks, but most models treat them separately. This paper analyzes the FJC model, deriving a conditional recursion and tractable approximations. Its simulations and community-structured experiments show bounded assimilation and sustained polarization under cascade-driven dynamics.
Problem
Online social networks couple information diffusion and opinion formation, while most existing models treat the processes separately.
Source: §I · p. 1Method
The paper derives an exact conditional layer-wise recursion for FJC, introduces a mean-field decoupling approximation and scalable tree-based reductions, and validates them through simulations.
Source: §I · pp. 1–2, E. Scalable Approximations · p. 8Results
In SBM networks with controlled initial opinions, FJC produces bounded assimilation and sustained polarization, unlike classical FJ’s typical global homogenization.
Source: §I · p. 2, §VI · p. 11Takeaways & Limitations
Cascade-driven interactions provide a theoretical foundation for understanding how opinions change in community-structured networks.
Source: §VI · p. 11
Abstract
from arXiv · showhide
Online social networks (OSNs) tightly couple information diffusion and opinion formation, yet most existing models treat these processes separately. Building on the Friedkin-Johnsen on Cascades (FJC) model, which integrates opinion updating within stochastic information cascades, we advance its analytical characterization and systematic validation. We derive an exact conditional layer-wise recursion that reveals the statistical coupling between diffusion realizations and opinion evolution, showing that a closed-form recursion in unconditional expectations is not available. To overcome this limitation, we introduce a mean-field decoupling approximation (MFA) that yields a tractable linear recursion for the expected opinion trajectory. We also develop scalable tree-based reductions that drastically reduce computational complexity while preserving structural diffusion properties, and validate the analytical recursion and its approximations through extensive simulations. Using Stochastic Block Model networks with controlled initial opinion configurations, we show that, unlike the classical Friedkin-Johnsen model, which typically leads to global homogenization, the FJC dynamics generate richer regimes of bounded assimilation and sustained polarization. Overall, this work bridges diffusion and influence modeling, providing a theoretical foundation for understanding how cascade-driven interactions reshape collective opinion landscapes in structured social networks.
I. INTRODUCTION
Online social networks couple information diffusion with opinion formation, motivating an analytical account of how cascade exposure shapes opinions. The paper situates this account within the FJ, cascade, and FJC literatures and describes the network and diffusion models it builds on.
I. INTRODUCTION
- The FJC model embeds Friedkin–Johnsen opinion updates in cascade exposure, but rigorous analytical characterization of cascade-driven opinion trajectories remained open.Each reshare transmits both content and the sharer’s current opinion, linking opinion change to diffusion paths.Source: §I · p. 1
II. RELATED WORK
- Prior work examined optimization and richer social mechanisms, while this study analyzes polarization in both opinion space and network topology.The paper distinguishes its analysis from work measuring only local disagreement or global divergence.Source: §II · p. 2
A. Network Representation
- The network is a directed graph whose edges indicate information flow, with a symmetric social matrix and a row-stochastic influence matrix that may be asymmetric.The influence matrix is obtained by normalizing interaction strengths.Source: A. Network Representation · p. 2
B. Information cascade
- Content propagation uses the Independent Cascade model, in which an active node may activate inactive neighbors with edge-specific probability.Source: B. Information cascade · p. 2
C. Friedkin-Johnsen model
- The classical FJ model combines agents’ initial predispositions with weighted neighbor opinions, while susceptibility controls openness to influence.Its synchronous dynamics converge to a unique equilibrium under the stated connectivity and anchoring conditions; the asynchronous variant used in FJC converges to the same equilibrium.Source: C. Friedkin-Johnsen model · p. 3
D. The Friedkin-Johnsen on Cascades Model
FJC embeds asynchronous FJ opinion updates in a stochastic cascade, so opinions change upon exposure and cascade participation determines which interactions occur. The model uses sequential posts and makes simplifying assumptions about exposure and propagation.
D. The Friedkin-Johnsen on Cascades Model
- A cascade determines when and between which nodes interactions occur, while the FJ update rule specifies how exposure changes opinions.The model preserves the FJ formulation while embedding it in short, burstlike online exposures.Source: D. The Friedkin-Johnsen on Cascades… · p. 3
- Each post spreads one at a time from a root through a cascade tree whose depth is bounded by the network diameter.Only nodes reached through the realized stochastic subtree participate; the tree sequence is extended to L layers for notational convenience without altering dynamics.Source: D. The Friedkin-Johnsen on Cascades… · p. 3
- Nodes become aware and update upon exposure from resharing predecessors, then reshare with probability θ_i to expose nodes in the next layer.A node processes its received predecessor opinions together in a single aggregated asynchronous FJ update.Source: D. The Friedkin-Johnsen on Cascades… · p. 3
- The model assumes opinion-neutral content, first-exposure tree-structured influence, and sequential rather than simultaneous cascades.The authors report validating the exposure and sequential-propagation simplifications against real Twitter cascade data in prior work, with high predictive accuracy for observed opinion dynamics.Source: D. The Friedkin-Johnsen on Cascades… · p. 4
IV. ANALYTICAL DERIVATION OF OPINION TRAJECTORIES
The analytical derivation organizes FJC opinion evolution by cascade layers. It replaces tracking every individual update time with layer-wise states and uses their sequential structure as the basis for linear state transitions.
IV. ANALYTICAL DERIVATION OF OPINION TRAJECTORIES
- The analysis formalizes FJC dynamics by combining layer progression, matrix-form opinion updates, cascade-layer probabilities, and approximations for tractable predictions.Source: §IV · p. 4
A. Sequential Opinion Updates
- For each post, nodes update sequentially layer by layer from the root outward, and the analysis tracks aggregated states rather than individual update times.Source: A. Sequential Opinion Updates · p. 4
- Each layer-wise state is the opinion vector after all nodes in that layer update; these states form sequential linear state transitions.Source: A. Sequential Opinion Updates · p. 4
B. Matrix Representation of the Update Process
The update process can be represented with linear transformations, first for an individual node and then for a full cascade layer. Because nodes in a layer update distinct coordinates, their effects can be aggregated at a common input state.
B. Matrix Representation of the Update Process
- Layer updates can be aggregated at the common input state because distinct nodes modify disjoint coordinates and their update operators commute.The resulting expression accounts for predecessor influence, non-updating nodes, and initial opinions.Source: B. Matrix Representation of the Update… · p. 4
- A node’s update is a local transformation of the current opinion vector, with all other components unchanged.Source: B. Matrix Representation of the Update… · p. 4
- The matrix form of a single update changes only the updating node’s component, combining predecessor influence, unchanged components, and initial-opinion contribution.The operators A_ij, B_i, and C_i encode these terms.Source: B. Matrix Representation of the Update… · p. 4
C. Probabilistic Structure of the Cascade
The section characterizes cascade membership through the resharing parents that can activate a node. It derives the probability of node activation and of each specific active-predecessor set.
C. Probabilistic Structure of the Cascade
- The parent set is partitioned into resharing parents φ(i), cascade members who do not reshare ξ(i), and parents outside the cascade ψ(i).Only resharing parents contribute to propagation.Source: C. Probabilistic Structure of the… · p. 5, Figure 2 · p. 5
- A node joins the cascade if at least one parent is in the cascade and reshares; δ_i denotes its probability of cascade membership.Nodes in layers 0 and 1 have δ_i = 1.Source: C. Probabilistic Structure of the… · p. 5
- For deeper layers, the probability that i joins the cascade is obtained by summing over subsets of parents outside the cascade and their non-resharing complements.The derivation assumes independence across nodes.Source: C. Probabilistic Structure of the… · p. 5
- The probability of a specific resharing-parent set A is computed by summing over configurations of parents outside the cascade among the remaining potential predecessors.Summing over nonempty A yields the probability δ_i that node i belongs to the cascade.Source: C. Probabilistic Structure of the… · pp. 5–6
D. Expected Opinion Dynamics
The section frames expected opinion dynamics as averaging opinion evolution over stochastic cascade realizations, rather than producing a deterministic final opinion vector.
D. Expected Opinion Dynamics
- Expected opinions are computed by averaging over possible cascade realizations because the stochastic FJC model does not yield a deterministic final opinion vector.The analysis combines a matrix update for a fixed cascade with probabilities describing possible cascade realizations.Source: D. Expected Opinion Dynamics · p. 6
1) Solution of a single cascade:
For a single cascade, opinion evolution across layers is statistically coupled to the cascade history, so an exact recursion in unconditional expectations is unavailable. The section introduces a mean-field decoupling approximation to obtain a tractable linear recursion for expected opinions.
1) Solution of a single cascade:
- An exact layer-wise recursion must condition on cascade history because the expected opinion of a predecessor averages over realizations inconsistent with its being an active resharing parent.Consequently, expected opinions across consecutive layers cannot be closed using unconditional expectations alone.Source: 1) Solution of a single cascade · p. 6
- The exact recursion characterizes the expected opinion at layer l conditional on the activation and resharing events through layer l−1.H_l−1 is the σ-algebra generated by those prior-layer events.Source: 1) Solution of a single cascade · p. 6
- The mean-field decoupling approximation neglects dependence between the resharing configuration and the preceding opinion vector, yielding a closed linear recursion for expected opinions.The recursion depends on marginal resharing probabilities and E[y_l−1].Source: 1) Solution of a single cascade · pp. 6–7
- In the layer-wise expectation calculation, nodes outside the cascade do not update, so their opinions remain at their initial values.The derivation separates nodes in the cascade from those outside it using cascade-membership indicators.Source: 1) Solution of a single cascade · p. 7
2) Solution of a sequence of cascades:
For a sequence of posts, the approximate single-cascade recursion is applied successively, carrying each cascade’s output forward as the next cascade’s initial condition.
2) Solution of a sequence of cascades:
- The expected final opinion after N cascades is computed by iteratively applying the MFA recursion, using each cascade’s output as the next initial condition.The procedure applies Theorem 1 for n = 1, …, N.Source: 2) Solution of a sequence of cascades · p. 8
E. Scalable Approximations
Tree-based reductions simplify cascade structure to lower computational costs, while retaining the most probable parent relations and, in one variant, pruning low-reachability nodes.
E. Scalable Approximations
- The exact framework evaluates all possible diffusion paths, but its matrices can be precomputed once for fixed r, Λ, and θ and reused across initial conditions.This supports repeated what-if evaluations despite the initial computational cost.Source: E. Scalable Approximations · p. 8
- The full probability tree (FT) attaches each reachable node to its most probable predecessor, leaving it with at most one parent.In the illustrated cascade, node 5 attaches to node 4 because θ4 > θ6.Source: E. Scalable Approximations · p. 8, Figure 3 · p. 8
- The tree is constructed from shortest weighted paths using edge weights −log(θi), with reachability probability given by the product of propagation probabilities along the path.The resulting directed acyclic tree encodes most-probable parent–child relations.Source: E. Scalable Approximations · p. 8
- Unlike FT, the thresholded tree (TT) removes nodes whose reachability probability falls below the user-defined threshold τ.Immediate neighbors of the cascade root remain included in both variants.Source: E. Scalable Approximations · p. 8, Figure 3 · p. 8
- Tree approximations drastically reduce the exponential combinations required by the exact method, enabling analysis of much larger networks.Each node has at most one possible predecessor, simplifying the combinatorial calculations.Source: E. Scalable Approximations · p. 8
V. PERFORMANCE EVALUATION
The evaluation uses a two-community stochastic block model and controlled susceptibility and initial-opinion settings to test approximation accuracy and cascade effects on polarization.
V. PERFORMANCE EVALUATION
- The evaluation uses a 100-node, two-block SBM with within- and between-community connection probabilities pin = 0.3 and pout = 0.01.The social matrix is the network adjacency matrix, and the influence matrix is derived by row normalization.Source: §V · p. 9, Figure 4 · p. 9
- The modular network supports controlled tests of how cross-community alignment shapes cascade influence, while its longer paths make it a demanding approximation benchmark.The between-community connection probability is much smaller than the within-community probability.Source: §V · p. 9
- Experiments vary resharing probability and susceptibility-centrality relationships, and compare FJC outcomes across four controlled initial opinion configurations.The configurations are polarized communities, neutral bridges, polarized bridges, and asymmetric polarization.Source: §V · p. 9
A. Validation of the approximations
The analytical predictions closely match simulated opinion distributions, while tree approximations preserve distributional trends but can miss aspects of their precise shape.
A. Validation of the approximations
- All distances between simulated distributions and theoretical results remain below 0.05.Simulation bars summarize 1,000 Monte Carlo runs with 95% confidence intervals.Source: A. Validation of the approximations · p. 9, Figure 5 · p. 10
- For FT and TT, Wasserstein-1 distances remain below 0.05, although Total Variation and Jensen–Shannon distances can be substantially larger.The approximations can capture overall distribution trends while mispredicting precise distribution shape.Source: A. Validation of the approximations · p. 9, Figure 5 · p. 10
- FT and TT have very similar accuracy, suggesting that pruning low-probability branches has limited effect under the tested settings.The results validate the analytical recursion and support MFA as a tractable surrogate for the exact stochastic dynamics.Source: A. Validation of the approximations · p. 10
B. Opinion dynamics and polarization
FJC restricts influence to realized cascade paths, unlike FJ’s global averaging, and can preserve polarization or produce bounded assimilation depending on community structure and susceptibility. The tested configurations illustrate how bridge susceptibility and cascade exposure shape these outcomes.
B. Opinion dynamics and polarization
- With opposing extreme communities, FJC preserves separation: peripheral nodes stay near their initial opinions while bridge nodes shift most.Centrality-proportional susceptibility intensifies this effect; inverse susceptibility makes bridges more stubborn and further reduces mixing.Source: B. Opinion dynamics and polarization · p. 10
- With one neutral and one extreme community, FJC weakens cross-community influence, so shifts remain limited and the initial separation largely persists.Overall shifts are smaller than in the polarized-community configuration.Source: B. Opinion dynamics and polarization · p. 10
- When bridges start neutral, their mediating role weakens under FJC: peripheral nodes shift more, and communities remain farther apart than under FJ.Stubborn bridges permit partial blending, whereas highly susceptible bridges maintain polarization.Source: B. Opinion dynamics and polarization · pp. 10–11
- With extreme-opinion bridges and neutral remaining nodes, the outcome depends strongly on bridge susceptibility and cascade exposure.At θ = 0.5, cascades reaching peripheral nodes may yield partial blending; sparse paths or effectively stubborn bridges preserve separation.Source: B. Opinion dynamics and polarization · p. 11
- FJ globally averages across social ties, whereas FJC restricts influence to realized cascade paths, preserving polarization or producing asymmetric blending.Across the studied SBM settings, FJC supports bounded assimilation and persistent polarization depending on community structure and susceptibility.Source: B. Opinion dynamics and polarization · p. 11, §VI · p. 11
VI. CONCLUSIONS
- The framework’s conclusions concern community-structured networks with controlled initial opinions, where cascade-mediated exposure reshapes collective opinion landscapes.The paper compares FJC and FJ on two-community SBM networks.Source: §VI · p. 11