Collective Influence in Networks

• Collective Influence is a network science concept that quantifies how local node interactions propagate to yield global connectivity through extended neighborhood measures.
• It employs adaptive heuristic removal to identify minimal node sets whose elimination efficiently dismantles the giant component compared to random or degree-based strategies.
• The concept extends to diverse applications including hypernetwork dismantling, opinion dynamics, and multiagent coordination, reinforcing its broad impact on complex systems.

Searching arXiv for the cited and closely related papers on collective influence. Collective influence is a family of concepts centered on how local entities, acting within a structured system, generate nonlocal or emergent effects at the collective scale. In network science, the term is most closely associated with a centrality measure introduced for optimal percolation, where the objective is to identify the smallest set of nodes whose removal destroys the giant component. In that setting, collective influence quantifies a node’s contribution to long-range connectivity by combining its reduced degree with the reduced degrees on a shell at distance $\ell$ (Morone et al., 2016, Kim et al., 2019). The same vocabulary has since been extended to higher-order networks, evolutionary games, Boolean-network stabilization, social influence and opinion dynamics, online emotional dynamics, and multiagent coordination, where “collective influence” denotes how local interactions aggregate into macroscopic patterns or control-relevant low-dimensional effects (Zhang et al., 2024, Szolnoki et al., 2016, Wang et al., 2017, Moussaid et al., 2013, Chmiel et al., 2011, Luo et al., 13 Jan 2026).

1. Origins in optimal percolation and network dismantling

In network science, collective influence arose from the optimal percolation problem: find the minimal set of nodes whose removal destroys the giant connected component. This problem is NP-hard, so practical approaches rely on heuristics that approximate which nodes are structurally most important for sustaining long-range connectivity (Morone et al., 2016). The central shift introduced by collective influence is that importance is not identified solely with degree or other one-node scores; rather, it is tied to how a node supports connectivity through its extended neighborhood.

The standard collective-influence centrality of node $i$ at radius $\ell$ is

$$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$

where $k_i$ is the degree of node $i$, $B(i,\ell)$ is the ball of radius $\ell$, and $\partial B(i,\ell)$ is the boundary at exact distance $\ell$ (Morone et al., 2016, Kim et al., 2019). The factor $i$0 counts non-backtracking branching from the source, while the shell sum measures branching opportunities at distance $i$1. This construction makes collective influence a local proxy for a global objective: reducing the non-backtracking eigenvalue that controls the existence of a giant component on locally tree-like networks (Morone et al., 2016).

A special case is $i$2, for which

$i$3

Because this is monotone in $i$4, ranking by $i$5 is equivalent to ranking by degree; high-degree adaptive percolation is therefore the $i$6 limit of collective-influence percolation (Kim et al., 2019). For $i$7, the ranking changes qualitatively: nodes can score highly not only because they are hubs, but because they connect shells of high-degree nodes and therefore act as bridges between structurally important regions (Kim et al., 2019).

The operational use of CI is adaptive. At each step one removes the node with the highest current CI score, updates the affected local region, and repeats. In the original scalable implementation, finite radius $i$8 and suitable heap-based bookkeeping yield overall complexity $i$9, which is the key reason the method is practical on massive graphs (Morone et al., 2016). More global variants were also introduced: CI propagation, which corresponds to $\ell$0 message passing, and CI belief propagation, an optimal-immunization variant. These improve performance only slightly—about 1–2% in the low-influencer tail—but increase complexity from $\ell$1 to $\ell$2, making them prohibitive for large-scale data (Morone et al., 2016).

2. Mathematical structure, adaptive attack, and criticality

Collective-influence percolation studies how the giant component disappears when nodes are removed adaptively according to CI scores. In the formulation analyzed on Erdős–Rényi networks with mean degree $\ell$3, the control parameter is the deactivated fraction $\ell$4, the order parameter is

$\ell$5

the probability that a randomly chosen node belongs to the giant component, and the susceptibility-like quantity is

$\ell$6

where the largest component is excluded from the sums (Kim et al., 2019). Near the transition, standard percolation theory predicts

$\ell$7

with mean-field values $\ell$8, $\ell$9, and $$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$0 on random graphs (Kim et al., 2019).

Finite-size scaling is expressed as

$$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$1

$$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$2

with corresponding critical scaling for $$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$3, $$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$4, and $$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$5 (Kim et al., 2019). Extensive Monte Carlo simulations up to $$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$6 together with an exact iterative numerical treatment for high-degree adaptive attacks show that CI-based dismantling is substantially more efficient than random removal in terms of the critical fraction $$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$7, but the transition remains continuous and mean-field (Kim et al., 2019).

For Erdős–Rényi networks with $$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$8, the reported critical points and exponents are:

Model	$$\mathrm{CI}_{\ell}(i) = (k_i - 1) \sum_{j \in \partial B(i,\ell)} (k_j - 1),$$9	$k_i$0	$k_i$1	$k_i$2
HDA (CI$k_i$3)	0.235 550	0.99(4)	1.00(4)	2.99(11)
CI$k_i$4	0.211 61(1)	1.02(4)	1.03(4)	3.06(13)
CI$k_i$5	0.206 01(1)	1.03(5)	1.03(4)	3.07(13)
Random percolation	0.714 285	1	1	3

These values show that CI dismantles the network at much smaller $k_i$6 than both random failures and degree-based adaptive attacks, yet the exponent estimates remain compatible with mean-field universality (Kim et al., 2019). The steep drop in $k_i$7 under CI is therefore visually abrupt but not discontinuous in the critical-exponent sense.

A notable structural feature is extensive degeneracy among top-centrality nodes. Near $k_i$8, there are $k_i$9 nodes sharing the same maximum degree or maximum CI value, with nonvanishing $i$0, so ties are resolved randomly among many equivalent candidates (Kim et al., 2019). This produces plateau structures in top-centrality values, saw-like patterns in the fraction of tied nodes, and cusp singularities in $i$1, $i$2, and

$i$3

without creating additional true critical points (Kim et al., 2019). This suggests that adaptive centrality-based attacks become effectively random-like near criticality, which may help explain why the universality class remains unchanged.

3. Algorithmic scaling and empirical validation of spreader selection

The same optimal-percolation logic that motivates node removal can be turned around to identify multiple spreaders. In this interpretation, the collectively influential set is the one whose positions minimize overlap and maximize the network-wide reach of a cascade, rather than merely maximizing individual centrality scores (Teng et al., 2016). The central claim is that influence is a many-body property: several high-degree nodes chosen independently can be redundant if they occupy the same dense region.

The CI framework was empirically evaluated on real information flows from APS, Facebook, Twitter, and LiveJournal by constructing virtual spreading processes based on observed tie strengths rather than relying on idealized epidemic models (Teng et al., 2016). For a spreader set $i$4 and node $i$5, the influence received from a source $i$6 along a path $i$7 is

$i$8

and the collective influence on node $i$9 is defined as

$B(i,\ell)$0

with global performance

$B(i,\ell)$1

for seed fraction $B(i,\ell)$2 (Teng et al., 2016). Across all four platforms and all tested $B(i,\ell)$3, CI-selected spreaders produced larger $B(i,\ell)$4 than degree, adaptive degree, PageRank, or $B(i,\ell)$5-core based heuristics (Teng et al., 2016).

This empirical result reinforces the original conceptual point: the number of connections or citations is not a deterministic indicator of collective importance. In the APS case, highly cited scientists were not necessarily those with highest collective influence in the coauthorship–citation structure, and on most platforms CI selected a mix of hubs and more modestly connected nodes occupying strategic positions (Teng et al., 2016). This suggests that network location relative to extended neighborhoods matters more than local popularity alone.

A closely related development extends the original percolation picture to general threshold cascades with first-order transitions. There, the decisive structures are not shells of high degree, but subcritical paths: chains of nodes that each need one more active neighbor to trigger (Pei et al., 2016). The resulting influence score, CI-TM$B(i,\ell)$6, is defined as the number of subcritical paths of length up to $B(i,\ell)$7 starting from node $B(i,\ell)$8, so a seed’s contribution is determined by how many such paths it can activate (Pei et al., 2016). This yields a linearly scalable algorithm for large threshold models, generalizing the CI idea from connectivity destruction to cascade initiation under peer pressure.

4. Extensions to higher-order and specialized dynamical systems

The original CI formalism was defined on simple graphs. Several later works generalize the idea to systems where pairwise edges are not the natural representation. One line of work addresses hypergraphs, where interactions occur through hyperedges connecting multiple nodes. In “HyperCI: A Higher Order Collective Influence Measure for Hypernetwork Dismantling” (Yan et al., 2021), the simple-graph formula is first lifted naively by replacing degree with hyperdegree,

$B(i,\ell)$9

and then refined into a three-factor measure

$\ell$0

which incorporates node co-occurrence, hyperedge expansion capacity, and the number of distinct nodes reachable through $\ell$1-hop hyperedges (Yan et al., 2021). On six real hypernetworks, HyperCI outperformed simple-network and hypernetwork baselines in accumulated normalized connectivity, indicating that higher-order structure changes which nodes are collectively decisive (Yan et al., 2021).

A different higher-order generalization appears in “Locating influential nodes in hypergraphs via fuzzy collective influence” (Zhang et al., 2024). There, influence is distributed over shells defined by $\ell$2-distance between hyperedges, but weighted continuously rather than by a hard frontier. The key ingredients are a radius

$\ell$3

a fuzzy membership

$\ell$4

and an entropy-weighted centrality

$\ell$5

averaged over $\ell$6 to obtain HDF and EHDF (Zhang et al., 2024). On six empirical hypergraphs, HDF and EHDF achieved the highest or second-highest Kendall $\ell$7 against SIR spreading influence and were particularly effective at identifying the top influential nodes (Zhang et al., 2024). This suggests that, in higher-order contagion, collective influence is more accurately captured by graded, entropy-weighted neighborhoods than by single-shell counts.

The CI concept has also been transferred to other dynamical systems. In evolutionary social dilemmas, a player’s strategy-passing potency is set proportional to

$\ell$8

so collective influence becomes a teaching-activity parameter in prisoner’s-dilemma dynamics (Szolnoki et al., 2016). The main finding is that there exists an optimal hierarchical depth $\ell$9 for promoting cooperation; degree-based influence is optimal only when temptation to defect is small, while $\partial B(i,\ell)$0 often performs best for stronger dilemmas (Szolnoki et al., 2016).

In Boolean networks, collective influence is adapted to stabilization rather than dismantling. The relevant operator is a modified non-backtracking matrix weighted by node sensitivities $\partial B(i,\ell)$1 and control variables $\partial B(i,\ell)$2, and the CI score becomes

$\partial B(i,\ell)$3

which approximates how controlling node $\partial B(i,\ell)$4 reduces the leading eigenvalue that governs damage spreading (Wang et al., 2017). On synthetic and real directed networks, this CI-based controller selection stabilized the system with fewer controlled nodes than high-degree, eigenvector-centrality, PageRank, or VoterRank heuristics (Wang et al., 2017).

5. Collective influence as information flow, emotion, and opinion dynamics

Outside classical network dismantling, several papers use “collective influence” to describe how local interactions produce emergent macroscopic states. In online emotional dynamics, the basic units are not nodes and shells but sequences of emotionally coded messages. Posts are assigned valence $\partial B(i,\ell)$5, emotional clusters are defined as consecutive runs of the same valence, and the hallmark of collective influence is that the probability a cluster continues grows with its length (Chmiel et al., 2011). The conditional probability

$\partial B(i,\ell)$6

for $\partial B(i,\ell)$7 implies preferential emotional attachment, and the corresponding cumulative cluster distribution

$\partial B(i,\ell)$8

fits long emotional runs much better than i.i.d. or simple Markov baselines (Chmiel et al., 2011). In BBC forums, average thread length increases with the absolute average emotional valence in the first ten comments, while emotional intensity decreases over the course of longer discussions, supporting the characterization of emotional expressiveness as “the fuel that sustains some e-communities” (Chmiel et al., 2011).

In controlled opinion-formation experiments, collective influence is grounded in empirically measured rules for how opinions and confidence change after exposure to another person’s estimate and confidence level (Moussaid et al., 2013). The update rule

$\partial B(i,\ell)$9

is modulated by normalized opinion distance and confidence difference, yielding three regimes: a confirmation zone, an influence zone, and a rejection zone (Moussaid et al., 2013). Simulations based on these measured micro-rules reveal two macroscopic attractors: an expert effect driven by a few highly confident individuals, and a majority effect driven by a critical mass of low-confidence individuals sharing similar views (Moussaid et al., 2013). A tipping point appears at roughly 15% experts, above which the expert attractor dominates the majority effect (Moussaid et al., 2013).

Opinion dynamics under media pressure provide another collective-influence formalism. In a well-mixed three-state model with opinions $\ell$0, $\ell$1, and undecided $\ell$2, the macroscopic variables $\ell$3, $\ell$4, and $\ell$5 obey

$\ell$6

where $\ell$7 is media bias toward $\ell$8, $\ell$9 summarize $i$00–$i$01 interactions, and $i$02 control persuasion of undecided agents (Colaiori et al., 2015). The resulting phase diagram contains four classes—finite critical mass, vanishing critical mass, zero critical mass, and total consensus—showing that media and social influence can reinforce or oppose each other, producing hysteresis and minority resilience (Colaiori et al., 2015).

Information-theoretic work further decomposes collective influence into modes of information flow. “Modes of Information Flow in Collective Cohesion” argues that transfer entropy and time-delayed mutual information conflate intrinsic, shared, and synergistic influences, even in two-particle leader–follower systems (Sattari et al., 2020). More recent work establishes a direct quantitative bridge between physically defined angular influence in a modified Vicsek model and transfer entropy. There, the pairwise influence $i$03 is the weighted angular contribution of neighbor $i$04 to the heading update of $i$05, and averaged pairwise influences serve as order parameters for flocking transitions (Pang et al., 25 Jun 2025). The study concludes that the partial-information decomposition based on intrinsic mutual information gives the most appropriate interpretation of influence in that system (Pang et al., 25 Jun 2025).

6. Contemporary reinterpretations in collective action and multiagent coordination

Recent work extends the term to settings where influence is not a graph-theoretic centrality but a low-dimensional latent quantity summarizing how many agents jointly affect a task-relevant state. In “Collective action through adaptive awareness” (Ariwayo et al., 6 Jul 2026), collective influence arises from networked opinion dynamics in which social reinforcement is shaped by an awareness-dependent nonlinearity: $i$06 Under a degree-based mean-field reduction this becomes

$i$07

Here awareness is encoded by the exponent $i$08, which alters the shape of the adoption nonlinearity rather than simply scaling rates (Ariwayo et al., 6 Jul 2026). The competition between effective social influence and abandonment produces discontinuous transitions, bistability, and hysteresis, so collective action can persist after the original environmental pressures weaken (Ariwayo et al., 6 Jul 2026). This suggests a broad reinterpretation of collective influence as a nonlinear responsiveness of populations to social reinforcement.

A more explicitly engineering-oriented use appears in multiagent reinforcement learning. “Scalable Multiagent Reinforcement Learning with Collective Influence Estimation” defines collective influence for agent $i$09 as the cumulative effect of all other agents on the task object, compressed into a low-dimensional latent variable

$i$10

where $i$11 is the state of the shared object and $i$12 is the Collective Influence Estimation Network (Luo et al., 13 Jan 2026). Instead of estimating each teammate’s action explicitly, the framework feeds $i$13 together with local state and object state into SAC critics and actors: $i$14

$i$15

Because the input and output dimensions of the estimator remain fixed as team size grows, the method avoids network expansion and allows new agents to be added without changing existing architectures (Luo et al., 13 Jan 2026). In three-arm cooperative lifting, decentralized CIEN-SAC reached near-centralized performance in 9 out of 10 runs under communication limits and transferred successfully to a real robotic platform (Luo et al., 13 Jan 2026). A plausible implication is that “collective influence” has become a design pattern for replacing explicit multiagent communication with object-centered latent interaction modeling.

A related but broader sociotechnical reinterpretation appears in “Dynamics of collective minds in online communities” (Ha et al., 10 Apr 2025). There the relevant state is a dynamic semantic network

$i$16

with topic frequencies $i$17 and topic similarities $i$18 (Ha et al., 10 Apr 2025). Editorial agenda-setting, amplification, reframing, trolling, counterspeech, and membership turnover alter the collective mind by perturbing these semantic-state variables and their update rules. Some interventions, such as alignment, are rapidly reversible, whereas amplification, reframing, trolling, and counterspeech generate persistent shifts in topic salience and semantic connections (Ha et al., 10 Apr 2025). This suggests that collective influence can also be understood as durable perturbation of shared representations rather than only as connectivity control or cascade triggering.

Across these diverse settings, the common thread is structural aggregation: collective influence refers to effects that cannot be reduced to isolated local importance. Whether expressed as shell-based centrality, subcritical-path counts, entropy-weighted higher-order neighborhoods, preferential emotional clustering, confidence-mediated opinion revision, awareness-shaped adoption nonlinearities, or latent object-centered interaction variables, the concept identifies how many local contributions combine into system-level outcomes (Morone et al., 2016, Pei et al., 2016, Zhang et al., 2024, Chmiel et al., 2011, Moussaid et al., 2013, Ariwayo et al., 6 Jul 2026, Luo et al., 13 Jan 2026).