This project was completed for the "Statistical Learning" course in the Data Science Master's program at Sapienza University of Rome. It involves implementing a Generalized Additive Model (GAM) using Local Polynomials and analyzing predictor importance through the Leave-One-Covariate-Out (LOCO) score.
Our approach scored 4th place out of 23 teams in the course's Kaggle competition, focused on a redshift prediction problem.
- Dataset: The training data consists of approximately 7,500 astronomical objects with redshift values and photometric color measurements (features: ug, gr, ri, iz, zy, and i). The goal is to predict the redshift of an additional 2,500 objects based on these photometric colors.
- Objective: Develop a GAM using local polynomial smoothers to model the relationship between redshift and photometric colors and assess predictor importance with LOCO.
- Competition: Redshift prediction problem on Kaggle, where the interpretability and predictive power of the model were key. This approach ranked 4th out of 23 teams.
