Using Random Forest to predict the presence of heart disease. See RandomForestClassifier documentation from scikit-learn.
With Random Forest, it is important to avoid overfitting and maximize out-of-sample accuracy by optimizing the model's hyperparameters. The following parameters are optimized in this example:
n_estimators
max_depth
max_features
The following optimization techniques are used:
- OOB error reduction
- grid search with 10-folds cross validation
The following feature importance techniques are used:
- Gini impurity
- Permutation feature importance
- Download Anaconda
- Download source code and install requirements
The heart disease dataset from Kaggle is utilized
Using the following parameters:
n_estimators = 90
max_depth = 2
max_featres = 'sqrt'
An in-sample F1 score of 88.81% and an out-of-sample F1 score of 88% are obtained. With any machine learning model, ideally you hope to obtain an in-sample accuracy a bit better than an out-of-sample accuracy. That way, you know your model is neither over nor underfit. Use these criteria, the modeling project was a success.