MLA-C01 Question 122
Single answerA retail company wants to predict customer churn using historical data. The dataset contains both numerical features (e.g., total purchase amount, visit frequency) and categorical features (e.g., customer segment, preferred payment method). The team has limited labeled data but expects to collect more over time. Which modeling approach is most appropriate to maximize initial performance and scalability as new data becomes available?
- A
Use a linear regression model to predict churn probability based on the numerical features.
- B
Train a decision tree model using all features and manually optimize hyperparameters.
- C
Develop an ensemble model such as a gradient-boosted tree algorithm that supports both numerical and categorical features.
- D
Implement a deep neural network with embeddings for categorical data and extensive hyperparameter tuning.
Show answer and explanation
Correct answer: C
Explanation
Gradient-boosted tree models are highly effective for tabular data with mixed feature types, and they are efficient even with limited labeled data. Additionally, they scale well as new data is added, making them a practical choice for predicting customer churn in this scenario. While deep learning models can also handle mixed feature types, they require substantial data and computational resources, which are not ideal in this case.
- A. Incorrect.
Linear regression is not suitable for a classification problem like churn prediction, as it is designed for regression tasks. Additionally, it cannot handle categorical features directly.
- B. Incorrect.
Decision trees may work but are prone to overfitting and are not scalable when more data becomes available. Manual hyperparameter optimization is also inefficient.
- C. Correct.
Gradient-boosted tree algorithms, such as XGBoost or AWS SageMaker's built-in XGBoost algorithm, are well-suited for datasets with mixed numerical and categorical features. They provide strong initial performance and scale effectively as new data is added.
- D. Incorrect.
Deep neural networks may work but require significantly more labeled data and computational resources to achieve good performance. This is not ideal given the limited labeled data available in this scenario.