Flu Shot Learning: Predict H1N1 and Seasonal Flu Vaccines

The aim of this project is to predict the probability of a subject taking the H1N1 and Seasonal flu vaccines according to the provided data. This project is built for the data science competition: https://www.drivendata.org/competitions/66/flu-shot-learning/.

In this readme, I will explain the steps I took to achieve my results in the competition:

• AUROC = 0.8442
• Top 14% of participants - as of writing this

Repository Structure

Sub-folders:

• input_data - raw data from competition;
• interim_data - preprocessed data to be used in modelling;
• output_data - model predictions to submit;
• models - pickled models to be imported by the notebooks.

Main folder:

• All the notebooks (.ipynb) and respective scripts (.py) for the project;
• requirements.txt - project dependencies;

Notebooks interpretation

1. EDA
2. PREPROCESSING
3. MODEL_SELECTION - performs cross-validation of models to select the best one;
4. TUNING - tunes the selected models hyperparameters, to improve score;
5. GENERAL - joins all the steps and performs predictions.

Solution Framework

In order to solve the problem I have applied the following Data Science mindset:

1. Explore the data using EDA - gain insight on the main aspects of the data such as distributions, trends, predictors, etc.
2. Clean data in PREPROCESSING - apply the gained insight to preprocess the data and getting it ready for model consumption.
3. Perform cross validation of MODELS - select the models I want to use; these models will be a basis to test different preprocessing assumptions and will eventually be part of the final model;
4. Tune some of the models using OPTUNA;
5. Get everything together and make predictions;
6. Iterate through every step applying different preprocessing assumptions, model building techniques and trying to optimize the model to the AUROC metric.