03 / Machine Learning / Regression
YouTube Channel Analytics
A regression workflow for predicting new subscribers from channel and video analytics, followed by ensemble and stacking experimentation.
Overview
A predictive-modelling project that uses channel and video analytics to estimate new subscribers, with later experimentation around model ensembles and stacking.
Problem
Subscriber growth is a numerical target influenced by multiple channel and video signals. The project explores how regression models can use these signals together.
Architecture
Analytics data feeds a regression workflow; predictions are assessed with complementary error and fit metrics.
- AnalyticsChannel / video data
- RegressionTree-based models
- SubscribersNumerical prediction
Implementation
- Use Python, Pandas and Scikit-learn for the modelling workflow.
- Explore Random Forest, CatBoost, XGBoost and LightGBM.
- Extend the experimentation to ensembles and stacking.
Challenges
A strong fit score does not describe every prediction error. Split design, leakage checks and the relationship between RMSE and MAE matter when interpreting subscriber predictions.
Technology
- Python
- Pandas
- Scikit-learn
- Random Forest
- CatBoost
- XGBoost
- LightGBM
Outcome
- R²
- ≈0.969
- RMSE
- ≈76.1
- MAE
- ≈27.0
Reported metrics are R² approximately 0.969, RMSE approximately 76.1 and MAE approximately 27.0. No additional improvement is claimed for the later stacking experiments.
The evaluation split, model associated with these scores and dataset details still need to be documented.
What I learned
Comparing complementary metrics gives a more useful picture than selecting a model on one score. Ensembles add another comparison to evaluate, rather than a guaranteed improvement.