Imitating Friends: Social Influence and Coordination

• Imitation of Friends is a social and computational process where local cues rather than global optimization direct action and coordination.
• It encompasses mechanisms like non-strategic copying, payoff-sensitive and selective imitation, applied in games, online choice, and recommendation systems.
• Mixed imitation strategies enhance consensus in networks, dynamically influence cooperation, and inform design in interactive robotics and social systems.

Imitation of friends denotes a class of social and computational processes in which action selection, preference formation, coordination, or inference is guided by socially local others rather than by global optimization. In the literature, “friends” are operationalized in several distinct ways: as network neighbors in coordination games, as anonymous counts of recommending contacts in online choice, as candidate or reliable contacts in recommendation systems, as same-role players in parallel groups of a repeated game, and as socially informative partners in interactive tasks (Vilone et al., 2012, Abbassi et al., 2011, Exman et al., 2014, Yu et al., 2019, Ueda, 22 Jul 2025). The resulting concept is therefore heterogeneous. It includes non-strategic copying, payoff-sensitive imitation, selective imitation based on inferred value alignment, and algorithmic reproduction of friend-recommendation behavior.

1. Operational meanings and conceptual scope

A first technical point is that the relevant literature does not treat friendship as a unitary psychological primitive. In the coordination model of social and strategic imitation, “friends” are simply network neighbors: agents copy or compare themselves only to adjacent nodes, and locality rather than tie strength is the defining property (Vilone et al., 2012). In the online-choice experiments on friend recommendations and crowd ratings, the social signal is not a named friend at all, but the number of friends who recommend or dislike an option; the studies therefore measure response to anonymous friend-count information rather than to specific trusted individuals (Abbassi et al., 2011). In the repeated-game analysis of imitation of friends, a “friend” is a player in the same role in another, parallel group, not a partner in the same interaction (Ueda, 22 Jul 2025). In recommendation systems, “friends” may be explicit contacts, “reliable friends” inferred from observed and unobserved social networks, or candidates generated by friend-of-a-friend structure and ranking heuristics (Exman et al., 2014, Yu et al., 2019).

This plurality matters because distinct formal objects are being imitated. What is copied may be a neighbor’s current action, a best-performing neighbor’s strategy, a friend-count cue, an action chosen by a same-role counterpart in another group, or an item consumed by a generated reliable friend. A common misconception is therefore to equate imitation of friends with simple conformity. The surveyed work shows instead that imitation can be non-strategic, strategically filtered, reward-alignment-sensitive, or embedded in recommendation and routing pipelines. This suggests that the shared abstraction is not intimacy but socially local information.

2. Local copying, consensus, and cooperation on networks

In coordination problems on networks, imitation of friends is formalized as local strategy updating. One influential model mixes two mechanisms: social imitation, represented by the voter model, and strategic imitation, represented by unconditional imitation. At each elementary time step, an agent updates socially with probability $q$ by copying a randomly chosen neighbor, and strategically with probability $1-q$ by copying the action of the best-performing neighbor if that neighbor earned more. Coordination is tracked by the density of active links $n_A(t)$, where $n_A(t)=0$ denotes full consensus (Vilone et al., 2012).

The main result is topology-dependent. On 1D and 2D lattices, pure strategic imitation freezes the system into disordered domains, whereas any nonzero amount of social imitation prevents freezing and leads to global consensus. On sparse complex networks, neither pure voter dynamics nor pure unconditional imitation is sufficient for efficient ordering, but their combination is: local fluctuations created by social imitation destabilize incompatible domains, while strategic imitation amplifies local majorities. The approach to consensus exhibits two regimes. When social imitation predominates, disagreement decays exponentially, $n_A(t)\sim \exp[-(1-q)t]$, with $\tau_q \sim \frac{1}{1-q}$. When strategic considerations dominate, the decay is algebraic, $n_A(t)\sim t^{-\gamma}$, with $\gamma \to 0$ as $q\to 0$, producing glassy slow ordering. The characteristic consensus time has a minimum at an intermediate $q^*$, so the fastest coordination occurs under a mixed imitation regime rather than under either pure rule (Vilone et al., 2012).

A related but broader framework studies imitation-based cooperation in dynamic social networks, where ties can be formed or broken while strategies coevolve. In CANDY, agents play a repeated Prisoner’s Dilemma on a graph $1-q$0, with payoff vector $1-q$1, equivalently $1-q$2. Strategy revision may follow imitate-payoff, conditional cooperation, or pairwise comparison, while rewiring may follow extreme popularity or active linking. The paper’s central structural result is the emergence of a cooperator-core and defector-periphery: dynamic partner selection isolates defectors and densifies cooperative clusters, thereby changing the local environment from which imitation operates (Bara et al., 2022).

This dynamic-network perspective clarifies a longstanding dispute about timescales. Some literatures report a threshold effect in the ratio $1-q$3, while others report a Goldilocks zone. The framework shows that both patterns can coexist under different assumptions about update rules and initial conditions. In this view, imitation of friends does not operate on a fixed substrate; rewiring changes who is available to imitate, and that coevolution can either promote or destroy cooperation (Bara et al., 2022).

3. Friend influence in online choice

In online choice, imitation of friends appears as a response to displayed social proof. A series of three Mechanical Turk studies asked participants to choose between two options while observing both crowd ratings and friend information. The core model is logistic: $1-q$4 Here $1-q$5 is the number of stars and $1-q$6 is the number of friend recommendations or negative friend opinions for option $1-q$7 (Abbassi et al., 2011).

Study	Friend effect	Star effect
Hotels, positive recommendations	$1-q$8, OR $1-q$9	$n_A(t)$0, OR $n_A(t)$1
Hotels, negative opinions	$n_A(t)$2, OR $n_A(t)$3	$n_A(t)$4, OR $n_A(t)$5
Movie trailers, positive recommendations	$n_A(t)$6, OR $n_A(t)$7	$n_A(t)$8, OR $n_A(t)$9

The quantitative pattern is consistent across studies. Friend signals are statistically significant and have the expected sign, so displayed friend information does shift stated choices. Yet aggregate ratings exert a larger marginal effect than friend counts. The strongest asymmetry concerns valence: negative opinions from friends are more influential than positive recommendations. The studies also report that choices become more random in lower-cost, lower-risk settings, which is reflected in the smaller coefficients for the movie-trailer task. Demographic covariates such as gender, age bracket, and education level were not statistically significant, and the paper interprets this as evidence that individuals trade off friend and crowd information in a similar fashion across these basic subgroups (Abbassi et al., 2011).

These findings limit any simple “friends dominate the crowd” thesis. The relevant marginal comparison is between one additional rating star and one additional friend signal, and the star effect is consistently larger. At the same time, the estimated friend effects are substantial rather than negligible. The paper therefore supports a view of imitation of friends as one input into a multi-source social-information calculus rather than as blind interpersonal copying.

4. Recommendation systems, friends-of-friends, and algorithmic mediation

A distinct research line studies not human imitation of friends but the imitation of friend-recommendation behavior itself. In “An Anti_Turing Test,” the target is to reproduce the composition, ordering, and temporal variation of social-network “People You May Know” lists from a reduced set of observable variables. The proposed RECOMM architecture combines friend-of-a-friend structure, randomization, and interestingness. The architectural modules are “Randomize inputs,” “Interestingness,” “Calculate Recommendation,” “Sorting & Threshold,” and “Decoration.” The ranking formulas are

$n_A(t)=0$0

and

$n_A(t)=0$1

Empirically, LinkedIn recommendations were observed to have network degree almost always “2nd,” indicating strong friend-of-a-friend structure, while randomization was invoked to explain list diversity, ordering changes, and visit-triggered refresh. The simulator is explicitly preliminary: it “does not yet reproduce faithfully actual recommendation lists, due to lack of precision of data gathered” (Exman et al., 2014).

This work is important because it treats friend recommendation as a modular black-box behavior rather than as a fully identified causal model. The strongest finding is that friend-of-a-friend is “clearly important,” but it is not sufficient on its own. Randomization and interestingness are necessary to reproduce observed list diversity and nontrivial distributional effects, including cases where a minority attribute becomes unexpectedly prominent among recommended candidates (Exman et al., 2014).

A different recommendation framework, RSGAN, begins from the opposite premise: explicit social links are often too sparse and too noisy to be trusted. It therefore seeks to generate “reliable friends” from both observed and unobserved social networks. The generator produces a distribution over candidate users, samples a user via Gumbel-Softmax, then samples one of that user’s consumed items; the discriminator is a BPR-style ranking model enforcing

$n_A(t)=0$2

A “generated friend” is not a synthetic identity but a sampled real user whose consumed items help predict the target user’s preferences. The model’s core claim is that reliable friends are dynamic and recommendation-specific, and need not coincide with explicit social ties (Yu et al., 2019).

The relation between direct friends and second-order social context also appears in social search. On the Gowalla network, efficient decentralized routing is achieved not by global knowledge but by the original distribution of friendship edges together with partial knowledge of friends of friends, denoted i-friends. The effect is strongly nonlinear: with optimal weights and $n_A(t)=0$3, search success rises from $n_A(t)=0$4 to $n_A(t)=0$5, while stretch improves from $n_A(t)=0$6 to $n_A(t)=0$7. Increasing $n_A(t)=0$8 beyond 15 “barely improves” success, even though the average number of distinct friends-of-friends per user is $n_A(t)=0$9 (Elsisy et al., 2019). This suggests that a shallow slice of second-order social information can capture much of the functional benefit of richer routing knowledge.

5. Selective imitation and the role of context

Imitation is not always indiscriminate. One experimental program argues that people preferentially imitate agents they infer to have similar reward functions. Participants were assigned one of two reward functions, $n_A(t)\sim \exp[-(1-q)t]$0 or $n_A(t)\sim \exp[-(1-q)t]$1, and observed demonstrators with the same or different reward functions. An aligned demonstrator is one with the same reward function as the participant; a misaligned demonstrator has the other reward function. Trajectory similarity is measured by

$n_A(t)\sim \exp[-(1-q)t]$2

and imitation is classified by choosing the demonstrator with maximal similarity (Taylor-Davies et al., 2023).

The strongest evidence appears in the path uncertainty phase, where participants overwhelmingly followed the aligned agent: Level 3 and Level 4 each yielded $n_A(t)\sim \exp[-(1-q)t]$3, with $n_A(t)\sim \exp[-(1-q)t]$4 and logistic accuracy $n_A(t)\sim \exp[-(1-q)t]$5. In goal generalization, aligned imitation persisted under uncertainty about new gem values: Level 5 yielded $n_A(t)\sim \exp[-(1-q)t]$6, $n_A(t)\sim \exp[-(1-q)t]$7, and Level 6 yielded $n_A(t)\sim \exp[-(1-q)t]$8, $n_A(t)\sim \exp[-(1-q)t]$9. Agent generalization was weaker but still present in one test level: Level 9 yielded $\tau_q \sim \frac{1}{1-q}$0, $\tau_q \sim \frac{1}{1-q}$1, whereas Level 10 was not significant. The paper does not manipulate friendship directly, but it proposes a mechanism by which friends or socially similar others may become privileged imitation targets: they are expected to have similar goals, and therefore their behavior is more informative for one’s own action choice (Taylor-Davies et al., 2023).

Another contextual effect concerns adversity. A general Bayesian model distinguishes decisions framed as identifying which option is best from decisions framed as estimating whether each option is good. In the symmetric case $\tau_q \sim \frac{1}{1-q}$2, the “best-option” ratio cancels $\tau_q \sim \frac{1}{1-q}$3 and is therefore independent of adversity: $\tau_q \sim \frac{1}{1-q}$4 By contrast, the “good-option” ratio depends on $\tau_q \sim \frac{1}{1-q}$5: $\tau_q \sim \frac{1}{1-q}$6 As $\tau_q \sim \frac{1}{1-q}$7 decreases, the relative influence of social information increases, predicting stronger imitation in adverse conditions. The paper reports support from killifish aggregation under negative odor conditions and from human reliance on another person’s opinion when the proportion of good cards is low (Pérez-Escudero et al., 2014).

Taken together, these works reject the idea that imitation of friends is a fixed propensity. It depends on inferred goal alignment, uncertainty, and environmental quality. This suggests that a friend signal is strongest when it is both informative and relevant to the imitator’s own payoff or objective structure.

6. Same-role imitation in repeated games and its formal limits

The most stringent formalization of imitation of friends appears in repeated games played in parallel. There are two identical games, and player $\tau_q \sim \frac{1}{1-q}$8 may imitate player $\tau_q \sim \frac{1}{1-q}$9, who occupies the same role in the other game. Tit-for-Tat is defined by

$n_A(t)\sim t^{-\gamma}$0

while Imitate-If-Better copies the friend only if the friend earned a higher payoff in the previous round. Unbeatability is defined pairwise: a strategy of player $n_A(t)\sim t^{-\gamma}$1 is unbeatable against player $n_A(t)\sim t^{-\gamma}$2 if

$n_A(t)\sim t^{-\gamma}$3

regardless of the behavior strategies of the other players (Ueda, 22 Jul 2025).

The main theoretical result is that imitation of friends is far more restrictive than imitation of opponents. Both naïve rules studied in the paper—TFT and IIB—are unbeatable if and only if the stage game is strongly payoff-monotonic for the relevant player: $n_A(t)\sim t^{-\gamma}$4 Under this condition, the player’s stage payoff depends only on the player’s own action and not on what the others do. The paper explicitly states that this condition is stronger than the existence condition for a fair zero-determinant strategy, which requires only weak payoff-monotonicity. When the strong condition holds, TFT becomes a fair ZD strategy and IIB becomes an unbeatable ZD strategy enforcing a nonnegative-payoff-gap condition. The paper also shows that $n_A(t)\sim t^{-\gamma}$5-IIB is unbeatable under weak payoff-monotonicity, so exploration weakens the required structural condition (Ueda, 22 Jul 2025).

The significance of these results is negative as much as positive. They show that simple cross-group imitation of a same-role friend is generally not robust in strategic environments, because the copied action is transplanted across groups with different co-player contexts. Unbeatability requires that the action’s payoff meaning be independent of that context. This sharply delimits the conditions under which imitation of friends can be guaranteed not to lose.

7. Interactive and embodied imitation

In robotics and human–robot interaction, imitation of a social partner is treated as an online coordination problem rather than as trajectory replay. A robot must infer the latent dynamics of a joint interaction, predict how the human partner will move next, and generate its own motion in real time. The formal model introduces partner-specific latent variables $n_A(t)\sim t^{-\gamma}$6 and $n_A(t)\sim t^{-\gamma}$7 for future motion windows and a shared latent variable $n_A(t)\sim t^{-\gamma}$8 for interaction dynamics common to both partners. Each partner’s recurrent state is

$n_A(t)\sim t^{-\gamma}$9

and the shared dynamics are inferred from either partner’s past (Bütepage et al., 2019).

The architecture has three parts: motion embedding for human and robot trajectories via VAEs, prediction of human-partner motion through a shared task-dynamics variable, and robot-trajectory generation conditioned on that variable. Training combines human–human interaction data with human–robot interaction demonstrations. The tasks are “hand-shake,” “hand-wave,” “parachute fist-bump,” and “rocket fist-bump,” collected using Rokoko suits and an ABB dual-armed YuMi IRB 14000. On robot test data, the Human Motion Embedding model achieves mean NRMSD $\gamma \to 0$0, compared with $\gamma \to 0$1 for Raw Data HR, $\gamma \to 0$2 for Raw Data R, and $\gamma \to 0$3 for the Gaussian model (Bütepage et al., 2019).

The technical lesson is that socially appropriate imitation is not fixed-trajectory mimicry. The correct action depends on timing, phase, onset, duration, and role. The paper’s shared-dynamics variable $\gamma \to 0$4 captures what both agents are jointly enacting rather than merely what one agent is doing. A plausible implication is that embodied imitation of a social partner is best understood as participation in a common latent interaction rather than direct copying of visible movement alone.