Flask Sockpuppet Detection App

Overview

This Flask web application detects potential sockpuppet accounts—deceptive online identities used to manipulate discussions—based on the textual content of user comments. The system uses machine learning with a focus on semantic embeddings and sentiment analysis to determine whether an input text likely originates from a sockpuppet account.

How It Works

1. Model & Metadata Initialization
   • Loads a pre-trained RandomForestClassifier model from model/random_forest.pkl.
   • Loads a list of input feature names from feature_metadata.pkl.
   • Optionally loads a lookup table (embeddings_lookup.csv) of precomputed Sentence-BERT embeddings for known inputs.
   • Loads a pre-trained PCA model (pca_model.pkl) for reducing embedding dimensionality.
2. Input Handling & Preprocessing
   • User input is captured from the web form (input_text) or as a JSON POST request.
   • Text undergoes cleaning, including:
     • Stripping URLs
     • Removing special characters
     • Collapsing extra whitespace
   • If the cleaned input is empty (e.g. just symbols or whitespace), an error message is returned.
3. Embedding Retrieval
   • If the input text exists in the precomputed embeddings lookup table, it reuses the stored vector.
   • Otherwise, it generates a fresh embedding using Sentence-BERT (paraphrase-MiniLM-L6-v2).
   • The embedding is:
     • Averaged across multiple runs (default: 3) for stability
     • Normalized to unit length
     • Transformed via PCA if available
4. Sentiment Feature Extraction
   • Using TextBlob, the following sentiment metrics are extracted:
     • Polarity (positive/negative tone)
     • Subjectivity (factual vs. opinion-based)
5. Feature Vector Assembly
   • The final feature vector is constructed to match the structure expected by the model:
     • Embedding components (e.g., embedding_0, embedding_1, ...)
     • Sentiment scores (polarity, subjectivity)
6. Prediction & Output
   • The assembled feature vector is passed into the RandomForestClassifier.
   • The model outputs probabilities for sockpuppet and non-sockpuppet.
   • The result and its confidence level are displayed on the result.html page and on the terminal.
7. Web Interface
   • A simple, intuitive Flask-based UI allows users to submit text.
   • The interface automatically launches in the default browser on app startup.

Getting Started

Prerequisites

Ensure the following dependencies are installed:

• Python 3.x: Required to run the application.
• Flask: Web framework used to serve the interface and handle routing.
• Pandas: Handles structured data, especially for embedding lookups and feature construction.
• NumPy: Supports numerical operations and vector transformations.
• TextBlob: Performs sentiment analysis, extracting polarity and subjectivity scores from input text.
• Scikit-learn: Powers the machine learning pipeline and PCA transformation.
• Joblib: Loads the pre-trained Random Forest and PCA models.
• Sentence-Transformers: Generates Sentence-BERT embeddings for input text.
• Pip: Required to install all dependencies.
• Standard Libraries: re, threading, and webbrowser (these come with Python and do not need separate installation).

Installation

1. Clone the repository:

   git clone https://github.com/zxyandreay/flask-sockpuppet-detection.git
   

2. Run the setup.bat script to set up the environment:

   @echo off
   cd /d %~dp0
   echo [INFO] Setting up virtual environment...
   
   if exist venv rmdir /s /q venv
   python -m venv venv
   call venv\Scripts\activate
   
   echo [INFO] Upgrading pip and installing dependencies...
   python -m pip install --upgrade pip
   pip install Flask pandas numpy textblob joblib scikit-learn sentence-transformers
   
   echo [SUCCESS] Setup complete!
   pause
   

Running the Application

1. Run the start.bat script to activate the virtual environment and start the Flask server:

   @echo off
   cd /d %~dp0
   call venv\Scripts\activate
   python app.py
   pause
   

2. The application will launch on http://127.0.0.1:5000/. Open a browser to interact with the web interface.

Features

• 🧼 Text Cleaning: Removes URLs, non-letter characters, and excess whitespace.
• 🧠 Dynamic Embeddings: Uses Sentence-BERT with optional averaging to encode input meaning.
• 💡 Sentiment Analysis: Adds interpretability via TextBlob’s polarity and subjectivity metrics.
• 📉 PCA Compression: Reduces dimensionality of embeddings for model efficiency.
• 🔍 Sockpuppet Detection: Predicts likelihood using a pre-trained Random Forest classifier.
• 🌐 User-Friendly Web UI: Flask-based frontend with automatic browser launch.
• ⚙️ Fallback Mechanisms: Embedding generation activates only if precomputed vectors are unavailable.

API Endpoints

Method	Route	Description
GET	/	Loads the main UI for user interaction
POST	/predict	Accepts user input and returns prediction

Technology Stack

• Backend: Python, Flask
• ML Model: Scikit-learn (Random Forest), Sentence-BERT, PCA
• NLP Tools: TextBlob (sentiment), Sentence-Transformers
• Frontend: HTML (served by Flask)
• Utilities: NumPy, Pandas, Joblib

License

This project is licensed under the MIT License, allowing free use, modification, and distribution, provided that credit is given to the original authors.

Note: This project was developed as part of an academic thesis requirement and is not intended for commercial use or large-scale deployment.