Kaggle Competition

As part of a 5-member coursework group project, I participated in a Kaggle competition on Elo Merchant Category Recommendation. The task was to accurately predict a customer's loyalty score for a given merchant using different Machine Learning models.Our team secured a top 3% ranking by leveraging an ensemble modeling approach.

My key contribution was spearheading the design and implementation of the innovative ensemble architecture using a two-phase blending technique. I strategically combined predictions from multiple specialized LightGBM models trained on different data subsets. In the first phase, I blended predictions from models trained with and without outliers using optimized weighting based on an adjustable probability threshold. In the second phase, I incorporated predictions from a dedicated outlier classification model. Through meticulous threshold tuning in both phases, I maximized predictive performance.
In addition, I engineered adjustments to handle low prediction values based on analysis of the training data distribution. This ensemble framework allowed our models' strengths to complement each other, leading to highly accurate and robust predictions

Collaborating closely with my talented teammates and pushing the boundaries of feature engineering were valuable experiences. Competing on a global Kaggle leaderboard taught me important lessons on real-world machine learning under tight time constraints. This project enhanced my data science skills and reaffirmed my passion for competitive challenges.