Mining and Intervention of Social Networks Information Cocoon Based on Multi-Layer Network Community Detection

TL;DR

This work tackles information cocoon formation in social networks driven by personalized recommendations by constructing a double-layer network that combines relational ties with feature-based similarity. It introduces two graph auto-encoder–based frameworks, Independent Graph Embedding (IGE-MTR) and Mixed Graph Embedding (MGE-MTR), to perform unsupervised multi-layer community detection via modularity-tensor reconstruction, and develops an intervention strategy based on influential users supported by a Markov-state propagation model. Empirical results on Cora, CiteSeer, and synthetic data show that both GAEs improve NMI and modularity measures, with MGE-MTR achieving the best performance, while the real-data study on Weibo demonstrates effective cocoon detection and attenuation of polarization through targeted intervention. The combination of multi-layer network modeling, modularity-based reconstruction, and intervention via influential nodes provides a scalable framework for monitoring information cocoons and guiding algorithmic regulation in online platforms.

Abstract

With the rapid development of information technology and the widespread utilization of recommendation algorithms, users are able to access information more conveniently, while the content they receive tends to be homogeneous. Homogeneous viewpoints and preferences tend to cluster users into sub-networks, leading to group polarization and increasing the likelihood of forming information cocoons. This paper aims to handle information cocoon phenomena in debates on social media. In order to investigate potential user connections, we construct a double-layer network that incorporates two dimensions: relational ties and feature-based similarity between users. Based on the structure of the multi-layer network, we promote two graph auto-encoder (GAE) based community detection algorithms, which can be applied to the partition and determination of information cocoons. This paper tests these two algorithms on Cora, Citeseer, and synthetic datasets, comparing them with existing multi-layer network unsupervised community detection algorithms. Numerical experiments illustrate that the algorithms proposed in this paper significantly improve prediction accuracy indicator NMI (normalized mutual information) and network topology indicator Q. Additionally, an influence-based intervention measure on which algorithms can operate is proposed. Through the Markov states transition model, we simulate the intervention effects, which illustrate that our community detection algorithms play a vital role in partitioning and determining information cocoons. Simultaneously, our intervention strategy alleviates the polarization of viewpoints and the formation of information cocoons with minimal intervention effort.

Figures (23)

• Figure 1: Browsing paths with or without a recommendation system
• Figure 2: The separation of Karate club network
• Figure 3: Research framework
• Figure 4: Multi-layer network community detection (IGE-MTR)
• Figure 5: Multi-layer network community detection (MGE-MTR)

Theorems & Definitions (8)

• Theorem 1
• Theorem 2
• proof : Proof of \ref{['optimal low rank reconstruction']}
• proof : Proof of \ref{['optimal low rank reconstruction2']}
• Lemma 1
• Lemma 2
• proof : Proof of \ref{['[Eckart and Young Theorem]']}
• proof : Proof of \ref{['low rank approximation']}