H-NeiFi: Non-Invasive and Consensus-Efficient Multi-Agent Opinion Guidance

Shijun Guo, Haoran Xu, Yaming Yang, Ziyu Guan, Wei Zhao, Xinyi Zhang, Yishan Song, Jiwei Chen

Abstract

The openness of social media enables the free exchange of opinions, but it also presents challenges in guiding opinion evolution towards global consensus. Existing methods often directly modify user views or enforce cross-group connections. These intrusive interventions undermine user autonomy, provoke psychological resistance, and reduce the efficiency of global consensus. Additionally, due to the lack of a long-term perspective, promoting local consensus often exacerbates divisions at the macro level. To address these issues, we propose the hierarchical, non-intrusive opinion guidance framework, H-NeiFi. It first establishes a two-layer dynamic model based on social roles, considering the behavioral characteristics of both experts and non-experts. Additionally, we introduce a non-intrusive neighbor filtering method that adaptively controls user communication channels. Using multi-agent reinforcement learning (MARL), we optimize information propagation paths through a long-term reward function, avoiding direct interference with user interactions. Experiments show that H-NeiFi increases consensus speed by 22.0% to 30.7% and maintains global convergence even in the absence of experts. This approach enables natural and efficient consensus guidance by protecting user interaction autonomy, offering a new paradigm for social network governance.

Introduction

Social media was intended to be a digital town square, yet it frequently devolves into a collection of soundproof rooms—the infamous "echo chambers." When platforms attempt to mitigate polarization by forcing users to encounter opposing views, the result is often counterproductive. This "backfire effect," driven by psychological reactance, causes users to entrench themselves further in their original positions.

The paper "H-NeiFi: Non-Invasive and Consensus-Efficient Multi-Agent Opinion Guidance" introduces a novel framework to address this. Instead of intrusive interventions like injecting bots or forcing unwelcome connections, H-NeiFi proposes a subtle, non-invasive method to lead social networks toward common ground by strategically filtering existing connections.

The H-NeiFi Framework

H-NeiFi (Hierarchical Neighbor Filtering) operates on a two-layer dynamic model that distinguishes between social roles:

  1. The Expert Layer (PCP): Users are categorized as experts or non-experts. Experts possess stable, "stubborn" opinions. H-NeiFi employs a Promotion Communication Pattern (PCP) to connect experts who are already somewhat aligned with the global goal, thereby accelerating the spread of that goal through established opinion leaders.
  2. The Non-Expert Layer (ACP): Non-experts are more flexible in their views. For this group, the system uses an Attention Communication Pattern (ACP) powered by Multi-Agent Reinforcement Learning (MARL).
Comparison of opinion guidance strategies: (a) initial distribution, (b) echo chamber effect, (c) intrusive baselines, and (d) the non-invasive H-NeiFi approach.

Non-Intrusive Neighbor Filtering

The core innovation of H-NeiFi is "neighbor filtering." Rather than adding new edges to the social graph—which users often perceive as intrusive—the system dynamically hides certain existing neighbors. The researchers argue that users are significantly less sensitive to the disappearance of a post than to the forced appearance of an unwelcome one. By pruning connections that lead to polarization, the system achieves consensus more naturally.

To manage this over the long term, MARL agents utilize a bidirectional LSTM network to process neighborhood data. They are trained with a dual-reward mechanism that balances moving users toward the global consensus while ensuring they do not become socially isolated. This prevents the network from fragmenting into disconnected clusters.

Experimental Evaluation

The researchers evaluated H-NeiFi using synthetic datasets simulating various social scenarios with 20 to 80 agents. Opinions were initialized using both uniform and uneven distributions to test robustness, following the standard "Bounded Confidence Model."

Key results include:

  • Consensus Speed: H-NeiFi increased the speed of reaching consensus by 22.0% to 30.7% compared to traditional baselines like Common-Neighbor (CNR) and Group-Pressure (GP).
  • Global Convergence: While baseline methods often resulted in multiple isolated opinion clusters, H-NeiFi consistently achieved a single global cluster.
  • Accuracy: The system achieved a significantly lower System Consensus Deviation (SCD), indicating that the final group opinion was much closer to the intended target goal.

Conclusion

H-NeiFi demonstrates that "subtraction" can be more effective than "addition" in social network governance. By respecting user autonomy and focusing on long-term connectivity, the framework provides a more efficient and less provocative path to consensus.

However, challenges remain regarding real-world validation and computational scalability for platforms with millions of users. Furthermore, the ethical implications of algorithmic filtering—even when non-invasive—require careful consideration regarding transparency and bias. Future work will likely explore these strategies within real-world trace data and existing recommendation ecosystems.

Summary updated

Abstract

The openness of social media enables the free exchange of opinions, but it also presents challenges in guiding opinion evolution towards global consensus. Existing methods often directly modify user views or enforce cross-group connections. These intrusive interventions undermine user autonomy, provoke psychological resistance, and reduce the efficiency of global consensus. Additionally, due to the lack of a long-term perspective, promoting local consensus often exacerbates divisions at the macro level. To address these issues, we propose the hierarchical, non-intrusive opinion guidance framework, H-NeiFi. It first establishes a two-layer dynamic model based on social roles, considering the behavioral characteristics of both experts and non-experts. Additionally, we introduce a non-intrusive neighbor filtering method that adaptively controls user communication channels. Using multi-agent reinforcement learning (MARL), we optimize information propagation paths through a long-term reward function, avoiding direct interference with user interactions. Experiments show that H-NeiFi increases consensus speed by 22.0% to 30.7% and maintains global convergence even in the absence of experts. This approach enables natural and efficient consensus guidance by protecting user interaction autonomy, offering a new paradigm for social network governance.
BibTeX
@inproceedings{guo2026h,
  author = {Shijun Guo and Haoran Xu and Yaming Yang and Ziyu Guan and Wei Zhao and Xinyi Zhang and Yishan Song and Jiwei Chen},
  title = {{H-NeiFi: Non-Invasive and Consensus-Efficient Multi-Agent Opinion Guidance}},
  booktitle = {The Web Conference},
  year = {2026},
  doi = {10.1145/3774904.3792113},
}
Figure