Google Professional Machine Learning Engineer Question 117
Single answerGoogle Cloud PlatformYou are building a machine learning model to predict customer churn for an online subscription service. Your dataset is stored in BigQuery and contains over 10 million rows of raw user data, including categorical and numerical features. You need to preprocess the data to remove missing values, normalize numerical features, and one-hot encode categorical features. Additionally, you aim to automate this preprocessing pipeline for scalability. Which approach should you take?
- A
Use TensorFlow Extended (TFX) to design a pipeline, leveraging TFX's Transform component for preprocessing tasks.
- B
Export the dataset from BigQuery to a CSV file, preprocess the data using a Python script, and re-import it into BigQuery.
- C
Use Dataflow to preprocess the data by creating a custom Apache Beam pipeline for missing value handling and feature engineering.
- D
Directly preprocess the data within BigQuery using SQL queries, followed by exporting the processed data for model training.
Show answer and explanation
Correct answer: A
Explanation
TFX is specifically designed for building end-to-end machine learning pipelines. Its Transform component allows you to perform preprocessing tasks in a scalable and reusable manner, and it integrates seamlessly with TensorFlow models. This makes it the most appropriate choice for preprocessing a large dataset stored in BigQuery for machine learning purposes.
- A. Correct.
This is the correct approach as TensorFlow Extended (TFX) provides a robust framework to design scalable and reusable pipelines for preprocessing tasks, including normalization, handling missing values, and one-hot encoding. The Transform component is particularly well-suited for these operations.
- B. Incorrect.
Exporting the dataset to a CSV file and using a Python script is not a scalable or efficient solution, especially for a dataset of this size. This approach lacks automation and would be difficult to maintain for future updates.
- C. Incorrect.
While Dataflow is a powerful tool for distributed data processing, creating a custom Apache Beam pipeline for this task would require significant manual effort and does not leverage the built-in preprocessing capabilities of TFX.
- D. Incorrect.
Processing data directly within BigQuery using SQL queries is feasible but not ideal for tasks like normalization and one-hot encoding, which are better handled by tools designed for machine learning workflows such as TFX.