This projects mainly focus on the insurance dataset. And, testing all sorts of machine learning models on the datasets. Datasets are obtained from Kaggle. Insurance company premium is mainly used to test all models.
- In this project, there are a total of 20 Machine Learning Models, including 4 Boosting models, 1 Stacking model, and 1 ANN(TensorFlow DNN) model.
- There are example notebooks from Kaggle. If you want to know further details and try something, you can read them.
- You can also see there is imbalanced target datasets and most Kaggle notebook train and test without oversampling.
- Therefore the accuracies of models are higher in testing with the normal dataset. But you will find there is overfit when predicting the minor class.
- Models are needed to adjust and hyperparameter tuning is needed.
-
Oversampling models take longer to train and results are not as good as normal dataset models.
-
Oversampling models are 30% more generalized than models that use the normal dataset.
-
Although oversampling models are low accuracy results, the confusion matrix shows a quite satisfactory result. Of course, models need to be tuned.
-
The normal dataset results are high, but the confusion matrix shows that the results of the minor class classification are only 60-70% accurate.
-
Random forest, Stacking, Xgboost and ANN models are quite promising, and you can fine-tune these models to get good results.
| Title | With SMOTE | Without SMOTE |
|---|---|---|
| All model f1 score |  |
 |
| All model f1 score plot |  |
 |
| logistic |  |
 |
| Bagging |  |
 |
| CatBoost |  |
 |
| Extra Tree |  |
 |
| Random forest Tree |  |
 |
| Stacking model |  |
 |
| ANN Matrix |  |
 |
| ANN Accuracy |  |
 |
| ANN Loss |  |
 |
- fine-tune the model to get high accuracy and reduce FP and FN.
- KFolding and Grid searching are necessary to know the dataset better.