Some binary classifiers are trained on a 2D dataset and visualised.
git clone https://github.com/PierreExeter/classifier_visualizations.gitadd conda env.
This visualisation by Scikit Learn compares the performance of various classifier. Adapted from here.
python 1_classifier_comparison.py
K-Nearest Neighbors is a popular classifier. After being fitted on the train data, it can predict a probability (between 0 and 1) for each point in the 2D feature space. The decision boundary is fixed at 0.5.
python 2_simple_plot_KNN.py
Changing the number of neighbors can have a significant effect on the test accuracy and may cause overfitting.
python 3_KNN_hyperparameter_tuning.py
The model is overfitting for low numbers of neighbors since the train accuracy is marginally higher than the test accuracy (i.e. the model does not generalise on unseen data). The best number of neighbors is 14 for this particular dataset.
python 3_SVC_hyperparameter_tuning.py
python 3_RFC_hyperparameter_tuning.py
Random forest is clearly overfitting the dataset.
add effect of dataset size
add ROC curve, F1 score, confusion matrix
add pipelines and cross validation and hyperparameter tuning