Which solution will improve recall in the LEAST amount of time?
Add class weights to the MLP’s loss function, and then retrain.
Gather more data by using Amazon Mechanical Turk, and then retrain.
Train a k-means algorithm instead of an MLP.
Train an anomaly detection model instead of an MLP.
Explanations:
Adding class weights to the MLP’s loss function can help to balance the importance of the target class relative to others, thereby improving recall without requiring extensive retraining or data collection. This approach directly addresses the issue of class imbalance in a quick and efficient manner.
Gathering more data through platforms like Amazon Mechanical Turk is often time-consuming and may require significant effort in data labeling and validation. This approach does not guarantee immediate improvement in recall, especially if the new data still reflects the original imbalance.
Training a k-means algorithm instead of an MLP does not inherently improve recall. K-means is primarily a clustering algorithm, not a classifier, and would not address the problem of class imbalance effectively, nor is it suitable for the task of multi-class classification.
Training an anomaly detection model may not be appropriate for the given task, as it typically focuses on identifying rare or outlier observations rather than improving recall for specific classes. This option would likely require more time for retraining and validation than adjusting the existing MLP with class weights.
To the best of my knowledge, the answer is:
Add class weights to the MLP’s loss function, and then retrain.
I weigh that the answer is:
Add class weights to the MLP’s loss function, and then retrain.