Predicting Airplane Aluminum Durability
Goal
Given a dataset of aluminum alloy properties and classes, we wanted to predict whether each alloy passed the extreme durability test.
We also wanted to achieve the highest score possible in the Kaggle competition that this problem was framed in.
Methods
We approached the problem first with a few models:
- logistic regression: classical baseline classification model; the model had a logistic cost of 0.51 and is interpretable, but its predictive power is somewhat lacking
- boosted trees: more robust in capturing the non-linear relationship between the response and the predictors; achieved logistic cost of 0.4264
These two models both have their own advantages, but they are not sufficient to meet our goals, so we engineered a stacked ensemble.
Final model and results
By individually tuning a CatBoost model, a XGB model, a Light GBM model and combining their output, the stacked ensemble model produces predictions that have the highest prediction accuracy, resulting in a 0.4216 logistic cost, which is a great improvement from our other models.
Even though we only ranked 15th, our score is very close to that of the first place, suggesting that our model performs almost as well as the best model.
What I did
- Data cleaning and Exploratory Data Analysis
- Automated cross-validation process through grid-searching combinations of best hyperparameters
- Trained and fine-tuned the boosted tree model
You can read more about our project through our report below