I’m working with an imbalanced dataset to predict strokes, where the positive class (stroke occurrence) is significantly underrepresented. Initially, I used logistic regression, but due to the class imbalance, I switched to a Random Forest model. After applying techniques such as random oversampling and adjusting the classification threshold, I've managed to improve my recall to approximately 61.3%. However, I am still facing a high false positive rate (178 instances) in my confusion matrix, which negatively impacts precision (17.6%). What additional strategies can I explore to further enhance precision while maintaining a good recall?
1
-
$\begingroup$ $1)$ Have you been looking at the whole PR curve? $//$ $2)$ Perhaps even better, consider the raw outputs and evaluate them directly without doing any classification at all. $//$ $3)$ Is there reason to believe your features will be extremely predictive of strokes with both a high recall and a high precision? It might be that you cannot reliably do much better without collecting additional features (which might not be available). A good analysis should honestly reflect such a shortcoming. $\endgroup$– DaveCommented Oct 29 at 14:10
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
You can generally try:
- using a metric that takes the imbalance into account
- class weights
- better algorithms (gbdts)
- tunning the prediction threshold or using continuous prediction directly
