From Local to Global Norm Emergence: Dissolving Self-reinforcing Substructures with Incremental Social Instruments

Yiwei Liu, Jiamou Liu, Kaibin Wan, Zhan Qin, Zijian Zhang, Bakhadyr Khoussainov, Liehuang ZhuICML 2021 · OpenReview

The Challenge of Local Norms

In decentralized systems—ranging from sensor networks to social media platforms—norms emerge through repeated interactions among agents. However, when a network is highly segregated into distinct communities, these groups often develop their own conflicting "local norms." These isolated pockets become self-reinforcing substructures (SRS): stable clusters of behavior that resist change and prevent the formation of a unified, global consensus.

Traditional solutions to this problem often involve drastically "rewiring" the network, which can destroy useful existing connections, or they require agents to possess unrealistically broad observation capabilities. To address this, researchers have proposed a more surgical approach: Incremental Social Instruments (ISI). The objective is to identify the absolute minimum number of new social ties required to break these local deadlocks and integrate the entire population into a single global norm.

Measuring Consensus: Entropy and the BA-Ratio

To model this phenomenon, the researchers treat norm emergence as a multi-agent system where individuals play repeated coordination games, earning rewards for aligning actions with their partners. To guide the strategic addition of new connections, they introduce several novel information-theoretic concepts:

  • A-entropy & B-entropy: "A-entropy" measures the uncertainty of interactions (how equally distributed interaction opportunities are), while "B-entropy" measures the uncertainty of beliefs (the diversity of opinions across the network).
  • The BA-ratio: This core metric is the ratio of B-entropy to A-entropy. A high BA-ratio indicates the presence of strong local norms (SRS), whereas a ratio of zero signifies that a global norm has been achieved.

Crucially, the authors mathematically establish the Small-Degree Principle. This principle proves that the most efficient way to lower the BA-ratio is not by connecting highly influential "hub" nodes, but rather by linking agents who have the fewest existing connections (low degree) and currently hold different beliefs.

Building on this, they developed the Adaptive ISI (AISI) algorithm. Instead of adding all new edges at once, AISI adds a single edge, allows agents to interact and update their beliefs via Q-learning, and then recalculates the optimal next edge based on the network's new state.

Experimental Validation and Key Insights

The researchers evaluated AISI across a mix of synthetic models (Gaussian Random Partition and LFR) and real-world network topologies, including email communications (Enron), co-authorship networks (DBLP), and human contact networks.

Low-Degree Nodes as Catalysts

Counter-intuitively, the study reveals that you do not need to link "influencers" to shift network behavior. Linking "outsiders" (low-degree nodes) is far more effective. It increases the overall interaction entropy more significantly, making the network functionally more connected with fewer structural changes.

Visualizing the Dissolution of Echo Chambers

The AISI algorithm consistently outperformed benchmark methods, including random edge addition and high-degree node linking. In experiments, the authors demonstrated that adding just a small number of strategic edges could completely dissolve multiple entrenched local norms into a single global consensus.

Visualizing the dissolution of local norms into a global consensus

Furthermore, the experiments validated the BA-ratio as a highly accurate proxy for social diversity. As the BA-ratio drops, the proportion of the population adopting the dominant norm increases predictably. The study also noted that networks with higher modularity (more distinct, tightly-knit communities) are inherently harder to unify, requiring more interventions to reach a global norm.

Limitations and Future Directions

While promising, the approach has limitations. The model assumes a "central organization"—such as a social media platform's recommendation engine—has the authority to suggest or create these new ties, which may not apply to purely organic systems. Additionally, the study relies on simple two-player coordination games, whereas real-world social norms often involve complex, multi-player dynamics. Future research will need to explore these complexities, as well as how "bad actors" might intentionally maintain local echo chambers to prevent global consensus.

Summary updated

Abstract

Norm emergence is a process where agents in a multi-agent system establish self-enforcing conformity through repeated interactions. When such interactions are confined to a social topology, several...
BibTeX
@inproceedings{liu2021from,
  author = {Yiwei Liu and Jiamou Liu and Kaibin Wan and Zhan Qin and Zijian Zhang and Bakhadyr Khoussainov and Liehuang Zhu},
  title = {{From Local to Global Norm Emergence: Dissolving Self-reinforcing Substructures with Incremental Social Instruments}},
  booktitle = {International Conference on Machine Learning},
  year = {2021},
  url = {https://openreview.net/forum?id=6x3ewY7-2D},
}
Figure