Three-Layer Structural Dynamic Governance Framework for Mastodon's Decentralization

An Empirical Study Based on Cross-Instance Interactions

📖 Overview

This project investigates the decentralized governance mechanism of Decentralized Online Social Networks (DOSNs) through a three-layer structural dynamic governance framework, using Mastodon as the benchmark platform.

By analyzing ~1.36 million cross-instance interaction records over 13 days, we reveal user behavior patterns, semantic community formation, and the optimal network structure for maintaining decentralization. The research provides data-driven insights and actionable tools for DOSN service management and governance.

🧐 Background & Motivation

Decentralized Online Social Networks (DOSNs) such as Mastodon, Bluesky, and Pleroma have emerged as alternatives to centralized platforms. For DOSNs, ecological health relies on two core pillars:

1. The frequency and quality of cross-instance interactions.
2. The decentralized balance of the overall network.

Despite the growing popularity of DOSNs, three key research gaps hinder effective governance:

• The quantitative impact of instance size on Cross-Instance Interaction Ratio (CIIR) remains unclear.
• The joint driving mechanism of language and topics in shaping cross-instance semantic communities lacks systematic analysis.
• A multi-dimensional evaluation system for measuring decentralization is absent.

This project aims to fill these gaps through empirical analysis and framework construction.

📂 Repository Structure

Directory	Description
docs/	Detailed research paper (PDF) and presentation slides.
data/	Standardized datasets, interaction matrices, and data dictionary.
src/preprocessing/	Scripts for data cleaning, event deduplication, and format standardization.
src/behavior_analysis/	(RQ1) Code for Instance Interaction Network Analysis.
src/semantic_analysis/	(RQ2) Louvain community detection and topic entropy analysis.
src/system_analysis/	(RQ3) Decentralization and Core Instance Analysis.

❓ Research Questions (RQs)

We focus on three interrelated research questions across three layers:

1. Behavioral Motivation Layer

RQ1: How does instance size affect cross-instance interaction behavior (measured by CIIR)?

2. Semantic Structure Layer

RQ2: How do languages and topics jointly shape the semantic community structure of DOSNs?

3. System Architecture Layer

RQ3: What is the optimal core instance scale window for maintaining a stable decentralized network structure?

📊 Data & Methodology

Data Source

The dataset is derived from the FediLive project, publicly available on Zenodo. It includes:

• 1,361,708 posts (livefeeds.json)
• Boost/favourite interaction data (boostersfavourites.json)
• Reply interaction data (reply.json)

Data Preprocessing

To ensure data quality, we performed three key steps:

1. Format Standardization: Unify IDs and timestamps.
2. Outlier & Noise Handling: Filter non-influential instances (Core instances: ≥20 active users).
3. Duplicate Removal: Generate unique event_id.

Core Datasets Generated

Dataset Name	Purpose
interaction_table.csv	Records all cross/intra-instance interactions (type, timestamp, parties)
instance_attributes.csv	Instance-level metrics (user counts, active users, top 5 topics)
interaction_matrix.csv	Interaction counts (reply/favourite/reblog) between instance pairs
instance_interaction_stat.csv	Detailed intra/cross-instance interaction statistics for each instance

🛠 Methodology & Analytical Pipeline

To investigate the three-layer governance framework, we employed a pipeline integrating statistical modeling, network science, and Natural Language Processing (NLP):

1. Behavioral Layer Analysis (RQ1)

• Metric Construction: We defined the Cross-Instance Interaction Rate (CIIR) as:

$$ \text{CIIR} = \frac{\text{Interaction}_{\text{external}}}{\text{Interaction}_{\text{total}}} $$

• Statistical Modeling: Used Ordinary Least Squares (OLS) Regression with quadratic terms and Generalized Additive Models (GAM) to capture non-linear relationships between instance size (log-transformed) and CIIR.
• Grouping & ANOVA: Categorized instances into Small, Medium, and Large based on active users to test for significant differences in interaction openness.

2. Semantic Structure Layer (RQ2)

• Community Detection: Applied the Louvain Algorithm to the instance interaction network to identify structural clusters.
• Homogeneity Measurement:
  • Utilized Information Entropy to measure the concentration of languages and topics within each community.
  • Calculated Jaccard Similarity to assess the overlap of interests between different linguistic groups.
• NLP Processing: Employed TF-IDF for keyword extraction from instance descriptions to define "Topic Profiles."

3. System Architecture Layer (RQ3)

• Multidimensional Centrality: Instead of simple degree centrality, we developed a composite index:
  • Output/Input Scores: Weighted out-degree/in-degree multiplied by the diversity of interaction partners.
  • Katz Centrality: Measuring the long-range influence of an instance across the entire federation.
  • Betweenness Centrality: Identifying "bridge" instances that control information flow.
• Concentration Analysis: Used the Gini Coefficient to quantify the inequality of influence distribution.
• Sensitivity Scanning: Conducted a K-point scan ($k=3$ to $30$) to observe how the Gini coefficient of core vs. peripheral networks fluctuates as the number of "core instances" increases.

---

💡 Detailed Key Findings

1. The "Inverted U-Shaped" Behavioral Law

We found that "Bigger is not always more open."

• Small Instances: Often remain isolated due to low activity.
• Medium Instances (Peak at ~20-50 active users): Exhibit the highest CIIR, serving as the "active ambassadors" of the federation.
• Large Instances: Show a "Centripetal Effect," where users tend to interact internally, creating a risk of "de facto centralization."

2. Dual-Track Semantic Structure: "Language as Border, Topic as Bridge"

• Linguistic Segregation: Language is the strongest predictor of community boundaries. Users naturally cluster with others speaking the same language, forming stable "cultural silos."
• Inter-community Connectivity: Interestingly, while languages divide, specific topics (e.g., Technology, Art, Gaming) act as bridges. A community might be linguistically homogeneous but topically diverse, allowing information to jump across language barriers through shared interests.

3. The "3–7 Core" Steady-State Window

Our structural analysis revealed a critical threshold for decentralized health:

• Optimal Stability: When the top 3 to 7 instances share the core influence, the network maintains a healthy balance between efficiency and decentralization.
• Centralization Reversion: Once the core expands beyond 7 instances or shrinks below 3, the Gini coefficient of the core network rises sharply, indicating that a few "super-nodes" are beginning to dominate the system, threatening the federated nature of the platform.

🚀 Contributions

Theoretical

• Proposed a three-layer structural dynamic governance framework (behavior-semantic-system).
• Revealed the "inverted U-shaped CIIR" and "language-topic dual-track" mechanisms.

Empirical

• Provided quantitative evidence from ~1.36 million real interactions.
• Validated the 3–7 core instance steady-state window.

Practical

• Offered actionable governance tools (CIIR, Gini coefficient) for DOSN operators.
• Proposed targeted strategies to mitigate centralization risks.

🛠️ Usage

1. Dataset Access

Download the original dataset from Zenodo or directly get standardized datasets used here for the project from the data/ folder .

2. Research Resources

Important

For the complete theoretical framework and detailed empirical results, please refer to the docs/ folder for the full paper and presentation slides.

3. Source Code & Analysis

All source code for experimental analysis is located in the /src directory, organized by research layer:

• Data Preprocessing (/src/preprocessing/): Scripts for data cleaning, format standardization, and event deduplication.
• Behavioral Analysis (/src/behavior_analysis/): Implementation of CIIR calculation, OLS regression, and GAM modeling for RQ1.
• Semantic Analysis (/src/semantic_analysis/): Louvain community detection and topic entropy calculations for RQ2.
• System Architecture (/src/system_analysis/): Multidimensional centrality (Katz, Betweenness) and Gini coefficient sensitivity scanning for RQ3.

To reproduce the findings, we recommend following the order: Preprocessing -> Analysis -> Visualization.

⚠️ Limitations & Future Work

Limitations:

• 13-day data window usually limits long-term evolution analysis.
• Semantic labels rely on instance descriptions.
• Lack of intervention experiments.

Future Work:

• Extend dataset to 6–12 months.
• Integrate user-level behavioral data.
• Conduct intervention experiments.
• Extend framework to Bluesky and Pleroma.