最も有効な Databricks-Machine-Learning-Professional 試験参考書

A Machine Learning Engineer has created a custom PyFunc model wrapper for a fraud detection system that needs to be registered in Unity Catalog under the schema risk_models in the production catalog. The model requires specific dependencies and must be accessible for governance and version control. The engineer is working on a dedicated cluster with Unity Catalog enabled, but the model registration is failing. Why is the model failing to be registered in this case?

• A. The model signature must be provided, as Unity Catalog requires all models to have signatures.
• B. The model artifacts must be stored in DBFS rather than Unity Catalog volumes.
• C. The PyFunc model must inherit from mlflow.sklearn.Model instead of mlflow.pyfunc.PythonModel.
• D. The cluster must be configured with shared access mode instead of dedicated mode.

What is the purpose of ParamGridBuilder?

• A. Feature storage
• B. Model deployment
• C. Creating hyperparameter combinations
• D. Tracking metrics

Which MLflow component is used to log parameters, metrics, and artifacts during model training?

• A. MLflow Projects
• B. MLflow Tracking
• C. Model Registry
• D. MLflow Models

A retail company wants to better forecast their sales of each SKU in every store in order to more accurately distribute their products. To achieve this, a Data Scientist proposes scaling their existing forecasting model to forecast individually for each combination of SKU and Store ID.
They have a cluster with 12 executors available in order to execute this. The current model is written using Pandas and the Prophet library for forecasting, with a Python function that receives a Pandas Data Frame with historic sales data as a parameter to train the forecasting model. For their next iteration, they want to improve the efficiency of this approach while using the least amount of effort and leveraging all available resources. Which approach will do this?

• A. 1. Use Spark repartition(12) to scale across all executors;
  2. Run to Pandas to convert the repartitioned DataFrame to Pandas so it can run on the original existing function;
  3. Add the Store ID and SKU to the Prophet model as exogenous variables.
• B. 1. Decorate the existing training function with the @pandas_udf;
  2. On the Spark DataFrame with all training data perform a GROUP BY by Store and SKU;
  3. Apply the decorated function to the grouped DataFrame with the apply function.
• C. 1. Rewrite the function using the Generalized Linear Regression models available in Spark MLlib, which allows parameterizing to reflect a similar behavior to that of Prophet;
  2. Perform a GROUP BY on the Spark DataFrame with the training data to group by Store and SKU;
  3. Apply the Spark UDF to the grouped DataFrame with the apply function.
• D. 1. Use a Spark query to identify all combinations of SKU and Store ID;
  2. Run a loop through every combination calling the existing Python function using Spark repartition(12) function so it distributes across all available executors.

A Data Scientist is building a propensity model for an e-commerce start-up. The company maintains 7GB of historical data and receives about 5MB of new transaction data daily. The goal is to generate daily purchase predictions for all users by 7:00 AM each morning. As the start-up is in its early stages, the data scientist must prioritize a highly cost-efficient approach. Which approach should the Data Scientist take?

• A. Use a single-node compute to build a model with libraries like scikit-learn. The model can then be integrated into an always-on streaming pipeline, to ensure immediate processing of incoming data in order to meet the SLA.
• B. Use a multi-node compute and leverage distributed frameworks like SparkML to train the model.
  The model can then be scheduled as a nightly batch job that runs on a multi-node compute.
• C. Use a single-node compute to build a model with libraries like scikit-learn. The model can then be deployed as an always-on REST API so that users can query the API to get the predictions whenever they want.
• D. Use a single-node memory compute to build a model with libraries like scikit-learn. The model can then be scheduled as a nightly batch job that runs on a single-node compute.