Geographical Structural Features of the WeChat Social Networks

Chuan Ai, Bin Chen, Hailiang Chen, Weihui Dai, Xiaogang QiuISPRS Int. J. Geo Inf 2020 · DOI

TL;DR

This paper introduces a novel statistical model, the Mixture of Dirichlet Network Distribution (MDND), to analyze the geographical structure of online social networks like WeChat. By modeling interactions between cities, MDND effectively captures how geographical proximity influences online social connections, offering a more interpretable explanation than traditional community detection methods.

What Problem Does This Paper Address?

Online social networks have become a rich source of data for understanding human interactions. A common observation is that users within the same "community" in these networks tend to be geographically clustered. While community detection algorithms can identify these clusters, they often lack an explanation for why this geographical aggregation occurs. Existing statistical models for networks typically focus on link probabilities without explicitly incorporating geographical information or providing a clear mechanism for this observed spatial correlation. This paper aims to bridge this gap by developing a statistical model that can both identify geographical patterns and offer a theoretical basis for their existence in online social networks.

Methods

The core of this paper is the development and application of the Mixture of Dirichlet Network Distribution (MDND) model. Unlike traditional community detection, which focuses on link density within communities, MDND models the network as a process of gradually generating links. It treats interactions between cities as the fundamental units, constructing "Spatio-info networks" where nodes represent cities and links represent information flow between them.

Key aspects of the MDND approach include:

  • Spatio-info Network Construction: User interaction data (e.g., browsing records) is geolocated using IP addresses to create city-level interaction networks.
  • Hierarchical Dirichlet Process: MDND is built upon a hierarchical Dirichlet process, allowing for flexible modeling of link distributions and their relationships to node classifications.
  • Gibbs Sampling: A collapsed Gibbs sampler is employed to estimate the model parameters, enabling the classification of both links and nodes.
  • Link and Node Classification: The model classifies links based on their characteristics and then infers node (city) classifications from the dominant link classes associated with them. This provides a more nuanced understanding than just node clustering.

The paper also compares MDND's performance against traditional community detection algorithms and administrative divisions using the Adjusted Rand Index (ARI) to quantify similarity.

Datasets

The study utilizes a dataset of WeChat user interaction records collected by "Fabonacci." Each record includes sharer ID, viewer ID, webpage ID, viewer's IP address, and browsing time. The IP addresses are geolocated using the Taobao IP Address Library to determine the geographical location of users. The data is processed to create city-level interaction networks, with three progressively larger networks (Network 1, 2, and 3) constructed for analysis.

Key Insights and Ideas

  • Geographical Proximity Matters Online: The study confirms that geographical proximity significantly influences online social network structures, with cities within the same province showing stronger connections.
  • MDND Explains Why: Unlike community detection, MDND provides a statistical framework that explains the geographical aggregation of nodes by linking it to the classification of links and their correlation with node attributes. This suggests that the type of information being shared is related to the location of the users sharing it.
  • Link Classification is Key: The ability of MDND to classify links, in addition to nodes, offers a richer understanding of network dynamics and information flow.
  • Model Performance Varies with Network Scale and Sparsity: MDND performs well on dense networks and shows increasing consistency with administrative divisions as iterations increase. However, its performance degrades significantly in sparse or unbalanced networks, highlighting limitations.

Key Results

  • High Agreement with Administrative Divisions: For dense networks (Network 1 and 2), MDND's node classification results show a high degree of agreement (ARI close to 1.00) with administrative divisions, comparable to or even exceeding traditional community detection methods.
  • Improved Separation in Complex Cases: In Network 3, which involves multiple provinces, MDND successfully separates cities from different provinces, a task where community detection showed some inconsistencies.
  • Limitations in Sparse/Unbalanced Networks: For a sparse network (Network 1 sample) and an unbalanced network (Guangdong vs. Guangxi), MDND's performance (ARI) was significantly lower and more volatile, indicating it struggles to extract meaningful geographical information in such scenarios. Community detection performed relatively better in these challenging cases.

Limitations and Future Work

  • Sparse and Unbalanced Networks: The paper explicitly acknowledges that MDND is not well-suited for extremely sparse or unbalanced Spatio-info networks. The authors suggest that new models are needed for these specific cases.
  • Empirical Validation: While MDND provides a theoretical basis for geographical aggregation, further empirical data is needed to confirm these hypotheses and refine the model.
  • Link Prediction: The authors propose that MDND's ability to model networks statistically makes it suitable for link prediction tasks, which could be valuable for applications like advertising delivery and Internet management.
  • Broader Applications: The study suggests that the link classification capabilities of MDND could be extended to classify information in social media networks, and the node grouping ability is comparable to community detection.

Summary updated

BibTeX
@article{ai2020geographical,
  author = {Chuan Ai and Bin Chen and Hailiang Chen and Weihui Dai and Xiaogang Qiu},
  title = {{Geographical Structural Features of the WeChat Social Networks}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2020},
  doi = {10.3390/ijgi9050290},
}
Figure