A machine learning project that predicts YouTube video success using metadata and channel characteristics.
Quick Facts
- Context: Term project for CS513 - Knowledge Discovery and Data Mining (Fall 2025)
- Tech Stack: Python, scikit-learn, Random Forest, SVM, MLP
- Links: GitHub Repo
Overview and Problem
The project processes over 85 million video records from the YouNiverse dataset to identify patterns that distinguish successful videos from others. The goal is to build a robust classification pipeline to predict video engagement accurately without data leakage.
What I Built
- Engineered an end-to-end ML pipeline from raw data ingestion to trained models.
- Implemented random and data preparation sampling with engagement filtering.
- Designed a feature engineering process with a strict train/test split and scaling to prevent data leakage (labels created post-split).
- Trained multiple classification models including Random Forest, Decision Tree, Linear SVC, K-Nearest Neighbors, and Multi-Layer Perceptron.
Key Results and Impact
- Optimized hyperparameters for 5 distinct models using GridSearchCV to maximize predictive performance.
- Processed a massive dataset of 85+ million records successfully.
Related: Projects MOC