Smoking Prediction using Synthetic Health Data

This project uses a synthetic health dataset to predict whether an individual smokes based on various health-related features. The dataset includes a range of biometric data such as age, weight, cholesterol levels, and more, and is designed for training machine learning models to make predictions about smoking habits.

Dataset

The dataset used in this project, which can be found here, consists of synthetic health records provided in CSV format:

• train.csv: The training dataset containing features and labels (smoking status).
• test.csv: The test dataset for making predictions.
• sample_submission.csv: A sample submission file for Kaggle competitions.

The columns in the dataset include:

• age, height(cm), weight(kg), waist(cm), eyesight(left), eyesight(right), hearing(left), hearing(right), systolic, relaxation, fasting blood sugar, cholesterol, triglyceride, HDL, LDL, hemoglobin, urine protein, serum creatinine, AST, ALT, GTP, dental caries, smoking

The target variable is the smoking column, where:

• 0: Non-smoker
• 1: Smoker

Project Goal

The goal of this project is to build a machine learning model that predicts the probability of smoking based on the health metrics provided in the dataset.

I used a HistGradientBoostingClassifier as the model and focused on optimizing its performance using cross-validation and hyperparameter tuning.

Approach

1. Exploratory Data Analysis (EDA):

   • Explored the dataset to identify missing values, data types, and the distribution of labels.
   • Identified and handled outliers by replacing them with boundary values.

2. Data Preprocessing:

   • No missing values were found in the dataset.
   • Features were selected and cleaned, focusing on continuous features for scaling.
   • Categorical features (like hearing and urine protein) were handled appropriately.

3. Modeling:

   • Built a pipeline with HistGradientBoostingClassifier and optimized it using cross-validation.
   • The model was tuned using GridSearchCV to find the best hyperparameters.

4. Evaluation:

   • The performance of the model was evaluated using ROC AUC as the metric, achieving an average score of 0.866 in cross-validation.

5. Final Submission:

   • The best model was trained on the entire training data and used to predict smoking probabilities for the test dataset.
   • My submission received a Kaggle competition score of 0.867, which is pretty good considering the best submission scored 0.879.

Conclusion

This project demonstrates the process of working with synthetic health data to build a predictive model for smoking habits. The dataset is well-structured, and using a tree-based model like HistGradientBoostingClassifier performed well despite the challenges of handling outliers and tuning the model.

Future Work Ideas

• Experimenting with different classifiers and feature engineering techniques.
• Incorporating more advanced feature selection.
• Testing with different validation strategies or resampling techniques.

Requirements

• Python 3.x
• pandas
• numpy
• matplotlib
• scikit-learn