Opinion Polarization Simulation on Social Networks

ISEF Category: Behavioral and Social Sciences

Subcategory: Social Psychology  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A small nudge from a recommender can change what people see next, and that can snowball fast. On a network, one extra echo can spread farther than you expect. You can model that spread and test when agreement turns into polarization.

What Is It?

This topic asks a simple but deep question: when a system keeps showing people content that matches what they already agree with, how fast do opinions split apart? In an agent-based simulation, each person becomes an agent, which means a simple rule-following character in your model. You then watch how those agents influence one another across a network, like students talking across different friend groups.

A small-world network is a map of connections that looks local but still has a few long-range links. That setup matters because real social networks work that way too. If your recommender boosts congruent content, meaning posts that match a person's current view, you can test whether the network stays mixed or breaks into separate camps.

Why This Is a Good Topic

This is a strong science fair topic because you can change one rule at a time and measure the result. You can test network structure, recommender strength, and initial opinion mix without needing a wet lab. The project connects to a real problem, how feeds and ranking systems shape what people believe, and it teaches simulation, data fitting, and social network analysis.

Research Questions

• How does recommender strength change the rate of opinion polarization?
• What is the effect of small-world rewiring probability on final opinion clustering?
• Does adding a few cross-cutting ties reduce echo chamber formation?
• To what extent do different starting opinion mixes change cascade size and spread speed?
• Which recommender rule best matches the shape of public Twitter cascade data?
• How does limiting congruent content amplification affect opinion diversity over time?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Python 3 installed.
• Mesa agent-based modeling package.
• NetworkX for network creation and analysis.
• pandas for data cleaning and summary tables.
• Jupyter Notebook or JupyterLab.
• Spreadsheet software for quick checks and charts.
• Public Twitter cascade dataset or a curated public repost network dataset.

Advanced Materials

• High-RAM workstation or university cluster access.
• Python environment with Mesa, NetworkX, pandas, SciPy, and statsmodels.
• Access to public Twitter cascade archives or a processed repost network dataset.
• Gephi for network visualization.
• Version control system such as Git.
• Statistical software for sensitivity analysis and model comparison.
• Documentation of dataset provenance and filtering rules.

Software & Tools

• Mesa: Builds and runs the agent-based simulation.
• Python: Controls the model logic and data analysis.
• NetworkX: Creates the small-world network and measures structure.
• pandas: Cleans cascade data and summarizes simulation output.
• Jupyter Notebook: Lets you explore model settings and chart results in one place.

Experiment Steps

1. Define the opinion state you will track and how agents can change it.
2. Build the small-world network and choose the network features you will compare.
3. Specify the recommender rule that boosts congruent content and set a no-recommender control.
4. Calibrate the model against public cascade data so your output matches real spread patterns.
5. Plan the metrics you will use, such as polarization, cluster separation, and cascade depth.
6. Run sensitivity checks to see which assumptions change the result the most.

Common Pitfalls

• Tuning the model on the same cascade data you use for evaluation, which makes the fit look better than it really is.
• Letting the recommender rule depend on hidden opinions you could not observe in the real data, which weakens the comparison.
• Treating one network layout as universal, which can hide how much structure drives the result.
• Using only average opinion change and missing polarization within subgroups, which can mask echo chambers.
• Forgetting to compare against a no-recommender baseline, which makes it hard to prove the amplifier changed anything.

What Makes This Competitive

A stronger version of this project goes beyond a single simulation run and tests whether the model still works across many real cascade patterns. You can compare different recommender rules, different network shapes, and different calibration targets, then report where each one fails. A careful sensitivity analysis and out-of-sample fit make the work much stronger than a simple demo. If you also explain which assumptions matter most, you turn the project into a real model study, not just a toy simulation.

Project Variations

• Use Reddit repost chains or quote-post networks instead of Twitter cascades to see whether the same polarization pattern appears.
• Swap the small-world network for a scale-free network and test whether hub-driven spread changes the outcome.
• Replace the congruent-content booster with a balanced feed rule and compare how fast polarization slows down.

Learn More

• Mesa documentation: Learn how to build agent-based models with Python, and find the official user guide online.
• NetworkX documentation: Review graph creation, small-world networks, and centrality measures on the project site.
• PubMed: Search for review articles on opinion dynamics, social influence, and algorithmic polarization.
• PNAS: Search the journal for studies on social contagion, echo chambers, and network diffusion.
• National Science Foundation public datasets: Look for social science data portals and network analysis resources.