PaperlayerView exact revision on arXiv

Source-linked AI summary

Social learning strategies modify the effect of network structure on group performance

Daniel Barkoczi, Mirta Galesic

arXiv:1606.00753v2cs.SIstat.AP

TL;DR

Research has found both efficient and inefficient networks to be superior, raising questions about how network structure and social learning strategy jointly shape group performance. This paper examines how those strategies interact with network structure and task complexity. It finds that the better-performing network type depends on copying strategy, while the strongest strategy also varies with task complexity.

  • Problem

    How social learning strategies interact with network structure to shape group performance remains a central question.

    Source: Network structure and social learning… · p. 4

  • Method

    The paper compares how social learning strategies and network structure jointly affect exploration, exploitation, and group performance.

    Source: Discussion · pp. 5–6

  • Results

    In complex tasks, efficient networks outperform inefficient networks under conformity, while the reverse holds under best-member copying; best-member strategies excel in simple tasks.

    Source: Discussion · pp. 4–5

  • Takeaways & Limitations

    Both well-connected and less well-connected networks can benefit performance, depending on the social learning strategy used.

    Source: Discussion · p. 5

    The study focuses on fixed network structures and groups relying on a single form of social learning.

    Source: Discussion · p. 6

Abstract

from arXiv · show

The structure of communication networks is an important determinant of the capacity of teams, organizations and societies to solve policy, business and science problems. Yet, previous studies reached contradictory results about the relationship between network structure and performance, finding support for the superiority of both well-connected efficient and poorly connected inefficient network structures. Here we argue that understanding how communication networks affect group performance requires taking into consideration the social learning strategies of individual team members. We show that efficient networks outperform inefficient networks when individuals rely on conformity by copying the most frequent solution among their contacts. However, inefficient networks are superior when individuals follow the best member by copying the group member with the highest payoff. In addition, groups relying on conformity based on a small sample of others excel at complex tasks, while groups following the best member achieve greatest performance for simple tasks. Our findings reconcile contradictory results in the literature and have broad implications for the study of social learning across disciplines.

Performance of different social learning strategies.

In a fully connected network, every strategy eventually reaches the global optimum in simple environments, but best-member strategies converge fastest.

Performance of different social learning strategies.

Performance in complex environments.

In complex environments, small-sample conformity achieves the highest long-run outcomes, while best-member strategies lead in the short run. Their different exploration and diffusion patterns help explain this contrast.

Performance in complex environments.

  • Small samples make frequency information noisier, increasing exploration and the chance of finding better solutions over time.Both small-sample strategies outperform their large-sample versions in the complex environment.Source: Performance in complex environments · p. 3
  • Small-sample conformity (s = 3) reaches the highest long-run outcomes, while best-member strategies achieve the highest short-run outcomes.Best-member copying quickly drives the population toward local optima where individual exploration cannot find better solutions.Source: Performance in complex environments · p. 4
  • Conformity converges more slowly and retains multiple solutions, allowing exploration while rare, superior options diffuse through the population.Its advantage stems from both heightened exploration and the diffusion of rare superior solutions, and the pattern recurs across complex landscapes.Source: Performance in complex environments · p. 4
  • Random copying exploits existing solutions but performs worse in the long run, while individual learning explores without spreading good solutions.The contrast illustrates the need to balance exploration and exploitation.Source: Performance in complex environments · p. 4

Network structure and social learning strategy interact.

Network performance depends on the social learning strategy: inefficient networks do better with best-member copying, while efficient networks do better with conformity. The findings help reconcile previously conflicting results.

Network structure and social learning strategy interact.

  • Inefficient networks outperform efficient ones under best-member copying, while efficient networks outperform inefficient ones under conformity.These patterns reproduce the opposing results reported in earlier studies, and conformity outperforms best-member copying in each network.Source: Network structure and social learning… · p. 4
  • Network diameter relates positively to average performance under best-member copying and negatively under conformity.Higher diameter indicates less efficient information spread; the authors note that some networks share diameter values, leaving an uneven number of values for analysis.Source: Network structure and social learning… · p. 4
  • The findings suggest performance is likely to rise when network structure and strategy promote opposite levels of exploration or exploitation.When both promote the same activity, performance is likely to drop.Source: Network structure and social learning… · p. 4

Discussion

Network advantages reverse with social-learning strategy, and the best strategy also depends on task complexity. The discussion uses these findings to reconcile prior results and notes the study’s fixed-network, single-strategy scope.

Discussion

  • Best-member strategies perform best on simple tasks, while weak conformity based on small samples yields the highest long-run outcomes on complex tasks.Best-member learning rapidly diffuses useful information toward local optima; slower conformity allows more exploration, and small samples help both strategies in complex environments.Source: Discussion · p. 4
  • Inefficient networks outperform efficient ones under the best-member strategy, whereas efficient networks outperform inefficient ones under conformity.Figure 3 reports this reversal across network categories; group performance therefore depends on both network structure and learning strategy.Source: Figure 3 · p. 5, Discussion · p. 5
  • The findings reproduce both prior patterns: inefficient-network superiority reported by Lazer and Friedman15 and efficient-network superiority reported by Mason and Watts16.The discussion attributes the different outcomes to the social-learning strategies used in the networks.Source: Discussion · p. 5
  • The study assumes fixed networks and a single social-learning form, leaving strategy mixtures and strategy selection for future work.The authors also propose studying real-world networks from large organizations or online platforms.Source: Discussion · p. 6
  • The results suggest that interventions changing the social environment without accounting for social-learning strategies might not produce the desired effects.The authors connect this implication to organizational learning, technological innovation, cultural evolution, and diffusion of innovations.Source: Discussion · p. 6

Generating the task environment.

The task environment is modeled as a tunably rugged NK landscape whose parameters determine solution interdependencies and payoff structure.

Generating the task environment.

  • NK landscapes vary in N solution components and K interdependencies, with binary solutions and element payoff contributions drawn uniformly from 0 to 1.A solution’s payoff is the average of its elements’ contributions.Source: Generating the task environment · p. 6
  • When K = 0, solution payoff is a simple average; when K > 0, each element’s contribution depends on K − 1 other interdependent elements.The interdependencies are assigned randomly, and changing one component can alter multiple payoff contributions.Source: Generating the task environment · p. 6
  • As K increases, single-component exploration becomes less effective because it changes the contributions of interdependent elements.With K = 0, single-component changes affect only that component’s contribution.Source: Generating the task environment · p. 6
  • Payoffs are normalized by the maximum landscape payoff, then raised to the eighth power to widen their distribution.The transformation makes most solutions mediocre and only a few very good.Source: Generating the task environment · pp. 6–7

Generating the networks.

The simulations compare networks with 100 nodes, holding degree at 19 except in the fully connected network, and generate additional structures through degree-preserving rewiring.

Generating the networks.

  • All networks have 100 nodes and degree 19, except the fully connected network, where each node has 99 neighbors.This degree-to-node ratio matches the cited Mason and Watts study.Source: Generating the networks · p. 7
  • The fully connected network links every node to all others, while the locally connected lattice links each node to its 19 closest consecutive neighbors.The lattice keeps degree fixed across nodes.Source: Generating the networks · p. 7
  • Eight additional networks are produced by degree-preserving rewiring that maximizes or minimizes closeness, betweenness, clustering, or network constraint.Rewiring runs for up to 1,000,000 iterations and is repeated 1,000 times to reduce trapping in local maxima or minima.Source: Generating the networks · p. 7

Simulation procedure.

The simulation applies social learning and, when it does not improve an agent’s payoff, individual learning across repeated runs of each strategy, task, and network combination.

Simulation procedure.

  • Each run starts with 100 agents assigned random solutions and repeats the learning procedure for 200 time steps.The random starting solutions provide high initial diversity.Source: Simulation procedure · p. 7
  • Agents search network contacts until sampling 3 or 9 solutions, then copy either the best-performing agent’s solution or the most frequent sampled solution.If sampled solutions are equally frequent, the agent uses individual learning.Source: Simulation procedure · p. 7
  • Agents adopt a socially identified solution only if its payoff is higher; otherwise, they test a one-digit modification and keep it only if payoff improves.Results are averaged over 1,000 repetitions for each strategy, task environment, and network structure.Source: Simulation procedure · p. 7

Additional information

This section provides publication, licensing, supplementary-material, and competing-interest information.

Additional information

  • Supplementary Information accompanies the paper.The passage directs readers to the journal’s Nature website.Source: Additional information · p. 7
  • The authors declare no competing financial interests.Source: Additional information · p. 8
  • The citation information identifies Barkoczi and Galesic’s article in Nature Communications 7, article 13109 (2016).Source: Additional information · p. 8
  • The paper is licensed under Creative Commons Attribution 4.0, subject to stated exceptions for third-party material.Permission may be required for material excluded from the license.Source: Additional information · p. 8

Related papers

Loading 1606.00753v2…