Which techniques should be used to meet these requirements?
Gather more data using Amazon Mechanical Turk and then retrain
Train an anomaly detection model instead of an MLP
Train an XGBoost model instead of an MLP
Add class weights to the MLP’s loss function and then retrain
Explanations:
Gathering more data using Amazon Mechanical Turk could help improve the model over time, but it is not a quick solution. In addition, the issue described is related to model performance (recall), not data scarcity.
Anomaly detection models are designed for rare or outlier events. This approach is not ideal for multiclass classification problems with an imbalanced target class, as it doesn’t directly address the recall issue in a typical classification setup.
XGBoost is a gradient boosting model that can be effective for certain types of classification problems, but it is not specifically tailored to improving recall in the context of an MLP model and imbalanced classes. This option would require more time and resources to implement compared to adjusting the MLP.
Adding class weights to the MLP’s loss function is a well-established technique for addressing class imbalance, especially when recall for a particular class is low. This change can help the model focus more on the underrepresented class and improve recall, making it a quick and effective solution.
I scheme that the answer is:
Add class weights to the MLP’s loss function and then retrain