johnnyyiu/poverty-prediction

[Advanced Regression] Predicting the Poverty Probability Index using socioeconomic data from 12600 individuals over 7 African countries # Poverty Prediction (Microsoft Data Science Professional Certification) ## Overview The motivation of creating this notebook is a data science competition hosted by **Microsoft** and **DrivenData**, which is a part of the Capstone project of the **Microsoft Professional Certificate in Data Science** and also my very first competition related to machine learning. By the end of the competition, I was at the **top 4%** on the leaderboard, and I am happy to share with you my approach, especially to those who are just beginning their journey on data science. ## Goal Participants are to predict the probability that individuals across 7 different countries live below the poverty line at the $2.50/day threshold, given other socioeconomic indicators. The probability of being in poverty was calculated using the **Poverty Probability Index (PPI)**, which estimates an individual's poverty status using 10 questions about a household’s characteristics and asset ownership. The remaining data comes from the **Financial Inclusion Insights** household surveys conducted by InterMedia. ## Data The dataset contains the PPI along with 58 features of 12,600 individuals across 7 different countries. ## Model I have hand-picked 3 regression-based models (Gboost, XGBoost and LightGBM) and used a 5-fold cross validation to evaluation their performance. A stacked model of the 3 is then tested for prediction results. The final r2 score was 0.4213, resulting a top 4% on the leaderboard. **The model stacking approach here is inspired by Serigne, make sure to check it out at:** Serigne's Stacked Regressions Notebook. Please check out my kaggle kernel for run result, and if you liked it - give it an up-vote! kaggle.com