Exam AWS Certified Machine Learning - Specialty topic 1 question 81 discussion

For me answer is D, adjust to higher weight for class of interest: https://androidkt.com/set-class-weight-for-imbalance-dataset-in-keras/. More data may/may not be available and a data labeling job will take time.

I believe is C, because we already made all changes possible in MLP hidden layers and the results have not improved then we must change model so XGBoot seems the best option

In this case, the data scientist is training a multilayer perceptron (MLP), which is a type of neural network, on a dataset with multiple classes. The target class of interest is unique compared to the other classes within the dataset, but it does not achieve an acceptable recall metric. Recall is a measure of how well the model can identify the relevant examples from the minority class. The data scientist has already tried varying the number and size of the MLP’s hidden layers, which has not significantly improved the results. A solution to improve recall must be implemented as quickly as possible.

The fastest one is D

"quickly as possible" mean do not change to new stuff, so it's D.

Not C, as the question ask for a quick solution. I accept D.

Answer C : https://towardsdatascience.com/boosting-techniques-in-python-predicting-hotel-cancellations-62b7a76ffa6c

Adding class weights to the MLP's loss function balances the class frequencies in the cost function during training, so the optimization process focuses more on the underrepresented class, improving recall.

I have done this before, class weights help with unbalanced data. Only logical solution that would help if not done, XGBoost could be different, but who knows, both NNs and XGBoost have comparable performance. Answer D!

In this example, it is necessary to improve recall as soon as possible, so instead of creating additional datasets, it is effective to change the weight of each class during learning.

C: 'distinct' indicates we can simplify this as a binary classification problem; then, NN is just overkill. plus, retraining a NN is much slower than training an XGboost model

I feel answer is B. Question says Target is different than the input data which is hint for anomaly detection.

I believe the answer is C because we need to use hyperparameters to improve model performance.

In case of the quickest possible way, D seems fine. For XGBoost, it will take a bit of time to code again

For me Answer A. Why no other model instead xgBoost, the model need more labeled data to be trained and learn more positive examples.