Instagram Political Bias Research 🔍

Academic research investigating political bias in social media algorithms through automated data collection and machine learning analysis.

📄 Full Paper: https://research.juleslemee.com

What This Research Does

This study examines how Instagram's algorithm creates political filter bubbles by systematically collecting and analyzing recommendation patterns across different user personas. In part 2, we also recreated a 2019 paper that created its own model to detect political bias in images.

Example of part 1: the scraper building a database of accounts with their political leaning

Part 1 methodology: automated user simulation and data collection

Example of part 2: the workflow to create the final classifier from the 2019 paper

Part 2 methodology: computer vision and algorithmic bias detection

Project Summary paper with findings and analysis

Why This Research Matters

The Problem: Social media algorithms create "echo chambers" that reinforce existing political beliefs, potentially polarizing society. Understanding how these systems work is crucial for digital literacy and informed civic participation.

Our Approach: By creating controlled experiments with bot accounts, we can systematically study algorithmic bias without relying on self-reported user data or proprietary company information.

Academic Impact: This research contributes to the growing field of algorithmic auditing and provides methodologies for studying recommendation systems across social platforms.

Research Components

Part 1: Automated Data Collection System

• Simulated User Behavior: Bot accounts with different political personas
• Systematic Interaction Tracking: Automated likes, follows, and engagement patterns
• Recommendation Analysis: Monitoring how algorithm suggestions change over time
• Political Content Classification: Building databases of politically-labeled accounts

Part 2: Computer Vision Analysis

• Bias Detection in Images: Adapting Thomas & Kovashka (2019) methodology
• High-Performance Computing: Custom SLURM implementations for GPU cluster training
• Feature Extraction: CLIP and Doc2Vec models for image analysis
• Classification Pipeline: End-to-end bias detection system

Key Features

✅ Automated Data Collection - Selenium-based Instagram scrapers
✅ Political Bias Detection - Computer vision analysis of images
✅ High-Performance Computing - SLURM GPU cluster implementation
✅ Academic Rigor - Peer-reviewed methodology and findings
✅ Reproducible Research - Complete codebase and documentation
✅ Real-World Impact - Insights into social media algorithmic bias

How to Reproduce This Research

Prerequisites

• Python 3.7+ with virtual environment capability
• Access to Instagram Accounts (for Part 1) or GPU cluster (for Part 2)
• Git for cloning repositories

Part 1: Data Collection Setup

1. Clone and Configure

   git clone https://github.com/juleslemee/instagram-algorithm-research.git
   cd instagram-algorithm-research/research-part-1
   pip install -r requirements.txt

2. Set Up Credentials

   cp input/keys/credentials.example.py input/keys/credentials.py
   cp input/keys/openai_key.example.py input/keys/openai_key.py
   # Edit files with your actual credentials

3. Run Data Collection

   python src/mainsim.py  # Automated user simulation
   python src/maincol.py  # Political account collection

Part 2: Machine Learning Analysis

1. Prepare Environment

   cd research-part-2
   pip install -r requirements.txt

2. Download Original Dataset

   • Visit Thomas & Kovashka project page
   • Download 130GB dataset to project directory

3. Run Analysis (GPU Cluster)

   module load slurm
   sbatch slurm_all.sh  # For continuous training

Project Structure

instagram-algorithm-research/
├── research-part-1/          # Instagram data collection
│   ├── src/                  # Scraping and automation scripts
│   ├── input/                # Configuration and credentials
│   └── output/               # Collected data and screenshots
└── research-part-2/          # Machine learning analysis
    ├── SLURM outputs/        # GPU cluster training logs
    ├── Models/               # Trained classifiers
    └── Features/             # Extracted image features

Research Findings

• Filter Bubble Formation: Documented rapid creation of political echo chambers
• Algorithmic Amplification: Measured bias in content recommendation systems
• Methodological Contributions: Novel approaches to social media algorithm auditing
• Technical Challenges: Identified limitations in automated data collection

Academic Context

Institution: McGill University
Course: COMP 396 - Individual Research Project
Supervisor: Professor Joseph Vybihal
Collaborator: Rayane Bouafia

Ethical Considerations

• All data collection followed academic ethical guidelines
• Bot accounts were clearly identified in profile descriptions
• No private user information was collected
• Research aims to promote algorithmic transparency

Technologies Used

• Web Automation: Selenium WebDriver, Python
• Machine Learning: PyTorch, CLIP, Doc2Vec, scikit-learn
• Computer Vision: OpenCV, PIL
• High-Performance Computing: SLURM, GPU clusters
• Data Processing: pandas, NumPy

Citations & References

This research builds upon and adapts:

• Thomas, C. & Kovashka, A. (2019). "Who's Behind the Camera? Identifying the Authorship of a Photograph"
• McGill University computing resources
• Instagram's public API and web interface

License

MIT License - See LICENSE file for academic and research use permissions.