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.