Human-Centric Community Detection in Hybrid Metaverse Networks with Integrated AI Entities

Shih-Hsuan Chiu, Ya-Wen Teng, De-Nian Yang, Ming-Syan Chen

TL;DR

As AI agents and digital twins become permanent fixtures in our social networks, traditional community detection methods struggle to prioritize human connections. This paper introduces the MetaCD problem and the CUSA algorithm, a novel framework that identifies human-centric communities by selectively retaining only the AI nodes that act as vital structural bridges for human interaction.

What Problem Does This Paper Address?

Traditional Social Network Analysis (SNA) assumes a world populated solely by humans. However, the rise of the Metaverse and Generative AI has birthed Hybrid Human-AI Social Networks (HASNs), where users interact with AI companions, virtual assistants, and digital twins.

Standard community detection aims to find "tightly-knit" groups, but applying it to HASNs creates a dilemma:

  1. AI-Complete Clustering: Treating AIs like humans leads to communities overrun by bots, making them useless for human-centric applications like targeted marketing or support groups.
  2. AI-Blind Clustering: Simply deleting all AI nodes can shatter the network's topology. In many cases, an AI entity might be the only "bridge" connecting two distinct human groups. Removing it results in fragmented, inaccurate communities.

The authors define this new challenge as MetaCD: finding clusters that maximize human closeness while minimizing AI presence, without sacrificing the structural integrity of the network.

Figure 2: Illustrating how different clustering strategies handle AI nodes. (d) shows the proposed AI-aware approach.

Methods

The authors propose CUSA (Customized AI-aware Simulated Annealing), which integrates three main components:

  • AI Scoring (Humanoid Score): Not all AI nodes are equal. CUSA assigns a "humanoid score" to AI entities based on their structural importance. It uses a weighted graph where human-human edges are prioritized ($hh > ha > aa$). The score combines three metrics: Eigenvector Centrality (influence), Betweenness Centrality (bridging), and Clustering Coefficient (redundancy).
  • AI-aware Louvain Clustering: The researchers modified the classic Louvain algorithm. Instead of just looking at structural modularity, the new objective function (HQ) includes a reward-penalty term based on the human-to-AI ratio within each cluster.
  • 3AC Framework: To avoid getting stuck in local optima (a common issue when deciding which nodes to prune), CUSA uses a Simulated Annealing-based "escape strategy." It iteratively removes the least "humanoid" AI nodes while checking if the overall human-centric modularity improves.
Figure 3: The iterative framework of the CUSA algorithm.

Datasets

Since real-world HASNs with ground-truth labels are scarce, the authors used a mix of synthetic and real data:

  • Reconfigured Benchmarks: Standard datasets (Cora, CiteSeer, PubMed) were transformed into HASNs using four novel synthesis strategies: k% random insertion, inverse degree-based insertion (AI helping lonely nodes), introverted/extroverted AI mix, and dual-personality AIs.
  • Timik: A real-world dataset from a social metaverse where digital twins act as AI nodes.
  • Scalability Benchmarks: Large-scale networks (YouTube and LiveJournal) were used to test the algorithm's performance on up to 1 million nodes.
Figure 4: The four proposed hybrid network (HASN) scenarios.

Key Insights and Ideas

  • AI as Social Glue: The most critical insight is that AI nodes shouldn't be treated as "anomalies" to be deleted. In hybrid networks, AIs often mimic human behavior and facilitate interactions that wouldn't otherwise happen.
  • The Bridge Effect: AI nodes with high betweenness centrality but low clustering coefficients are often vital "bridges." Retaining them actually increases the detected closeness between human members.
  • Human-Centric Modularity (HQ): By baking the human-to-AI ratio directly into the optimization objective, the algorithm naturally gravitates toward communities that are useful for human engagement rather than just mathematically dense groups of nodes.

Key Results

  • Superior Accuracy: CUSA consistently outperformed baselines (Spectral, Louvain, and GNN-based methods like GCC and LSEnet). On the Cora dataset, CUSA achieved a 17.6% relative gain in human-centric modularity (HQ) compared to standard Louvain.
  • Robustness to AI Density: In the Timik dataset, where AI nodes make up 46% of the network, traditional "all-removal" strategies failed because the remaining human network was too sparse. CUSA maintained high performance by identifying the specific AIs holding the network together.
  • Scalability: While the theoretical complexity is higher, empirical tests show CUSA scales near-linearly ($O(|E|)$) with the number of edges, making it practical for real-world social platforms.

Limitations and Future Work

  • Structural Focus: The current model relies entirely on graph topology. The authors note that future iterations could incorporate semantic features, such as the actual content of human-AI chats or shared interests, to further refine clusters.
  • Node Identification: CUSA assumes we already know which nodes are AI. While this is true for platform owners, a more general version might need to integrate AI-detection (Graph Anomaly Detection) as a preprocessing step.
  • Dynamic Networks: While the paper touches on network evolution, more work is needed to handle real-time, streaming HASNs where nodes and edges appear and disappear rapidly.

Summary updated

Abstract

Community detection is a cornerstone problem in social network analysis (SNA), aimed at identifying cohesive communities with minimal external links. However, the rise of generative AI and Metaverse introduce complexities by creating hybrid human-AI social networks (denoted by HASNs), where traditional methods fall short, especially in human-centric settings. This paper introduces a novel community detection problem in HASNs (denoted by MetaCD), which seeks to enhance human connectivity within communities while reducing the presence of AI nodes. Effective processing of MetaCD poses challenges due to the delicate trade-off between excluding certain AI nodes and maintaining community structure. To address this, we propose CUSA, an innovative framework incorporating AI-aware clustering techniques that navigate this trade-off by selectively retaining AI nodes that contribute to community integrity. Furthermore, given the scarcity of real-world HASNs, we devise four strategies for synthesizing these networks under various hypothetical scenarios. Empirical evaluations on real social networks, reconfigured as HASNs, demonstrate the effectiveness and practicality of our approach compared to traditional non-deep learning and graph neural network (GNN)-based methods.
BibTeX
@inproceedings{chiu2025human,
  author = {Shih-Hsuan Chiu and Ya-Wen Teng and De-Nian Yang and Ming-Syan Chen},
  title = {{Human-Centric Community Detection in Hybrid Metaverse Networks with Integrated AI Entities}},
  booktitle = {The Web Conference},
  year = {2025},
}
Figure