MLA-C01 Question 120
Single answerA financial services company wants to predict customer churn and has a dataset with categorical, numerical, and timestamp features. The dataset contains missing values and imbalanced target classes. The team has decided to use AWS SageMaker to train the model. Which modeling approach should you choose to address both the imbalanced dataset and the feature types effectively?
- A
Use SageMaker Linear Learner and enable the 'balance_data' hyperparameter
- B
Use SageMaker XGBoost and preprocess the dataset to handle missing values and imbalance
- C
Use SageMaker K-Means for clustering the data and then train a model on the clusters
- D
Use SageMaker DeepAR for time-series forecasting to predict churn
- E
Use SageMaker BlazingText as it supports categorical and numerical features natively
Show answer and explanation
Correct answer: B
Explanation
The most suitable modeling approach in this scenario is to use SageMaker XGBoost. It is a robust algorithm for classification tasks and supports handling categorical, numerical, and missing values directly without extensive manual preprocessing. Additionally, it offers hyperparameters like scale_pos_weight to address imbalanced datasets, making it the ideal choice for predicting customer churn in this case.
- A. Incorrect.
Linear Learner is a good algorithm for binary classification, but it requires careful handling of missing values in preprocessing. While the 'balance_data' hyperparameter can help with class imbalance, it does not address the diverse feature types in the dataset.
- B. Correct.
XGBoost is well-suited for this scenario because it can handle categorical, numerical, and missing values natively. It also works well with imbalanced datasets when combined with techniques like adjusting the scale_pos_weight parameter.
- C. Incorrect.
K-Means is a clustering algorithm and is not suitable for supervised binary classification tasks such as predicting customer churn.
- D. Incorrect.
DeepAR is specifically designed for time-series forecasting and not for binary classification tasks like churn prediction.
- E. Incorrect.
BlazingText is optimized for text data and is not appropriate for datasets with mixed feature types like numerical, categorical, and timestamps.