Google Professional Data Engineer Question 215
Select 3Google Cloud PlatformYou are working on a machine learning project to predict customer churn for a subscription-based service. The data contains features such as customer demographics, subscription history, and activity logs. You need to prepare the data for feature engineering to ensure consistency between the training and serving phases. Which steps should you take?
- A
Normalize numerical features to have a mean of 0 and a standard deviation of 1.
- B
One-hot encode categorical features based on unique values observed in the training data.
- C
Identify and remove outliers from the test dataset to improve model performance.
- D
Use a transformation pipeline to ensure the same preprocessing steps apply to both training and serving data.
- E
Randomly shuffle the features in the dataset to avoid data leakage.
Show answer and explanation
Correct answers: A, B, D
Explanation
Preparing data for feature engineering involves ensuring consistency between training and serving phases. This includes scaling numerical features, encoding categorical features consistently, and using a unified pipeline for transformations. These steps prevent discrepancies that could degrade model performance in production.
- A. Correct.
Correct. Normalizing numerical features ensures they are scaled consistently, which is critical for many machine learning models such as logistic regression and neural networks.
- B. Correct.
Correct. One-hot encoding ensures categorical features are represented numerically, and basing this on training data ensures consistency during both training and serving.
- C. Incorrect.
Incorrect. Outlier removal should generally be handled during training data preprocessing, not directly on the test dataset, as the test data should represent real-world scenarios.
- D. Correct.
Correct. Using a transformation pipeline ensures that the same preprocessing logic is applied to both training and serving, avoiding inconsistencies.
- E. Incorrect.
Incorrect. Randomly shuffling features would disrupt the data structure and relationships, leading to incorrect model training and evaluation.