Snowflake SnowPro Advanced: Data Scientist Certification : DSA-C03 exam

You are deploying a machine learning model to Snowflake using a Python UDF. The model predicts customer churn based on a set of features. You need to handle missing values in the input data'. Which of the following methods is the MOST efficient and robust way to handle missing values within the UDF, assuming performance is critical and you don't want to modify the underlying data tables?

• A. Pre-process the data in Snowflake using SQL queries to replace missing values with the mean for numerical features and the mode for categorical features before calling the UDF.
• B. Raise an exception within the UDF when a missing value is encountered, forcing the calling application to handle the missing values.
• C. Use within the UDF, replacing missing values with a global constant (e.g., 0) defined outside the UDF. This constant is pre-calculated based on the training dataset's missing value distribution.
• D. Use within the UDF to forward fill missing values. This assumes the data is ordered in a meaningful way, allowing for reasonable imputation.
• E. Implement a custom imputation strategy using 'numpy.where' within the UDF, basing the imputation value on a weighted average of other features in the row.

You've trained a sentiment analysis model in Snowflake using Snowpark Python and deployed it as a UDF. After several weeks, you notice the model's performance has degraded significantly. You suspect concept drift. Which of the following actions represent the MOST effective and comprehensive approach to address this situation, considering the entire Machine Learning Lifecycle, including monitoring, retraining, and model versioning? Assume you have monitoring in place that alerted you to the drift.

• A. Retrain the model on a sample of the most recent data, overwriting the original model files in your Snowflake stage and updating the UDF definition. Keep no record of the old model.
• B. Adjust the existing model's parameters manually to compensate for the observed performance degradation without retraining or versioning.
• C. Disable the model and revert to a rule-based system, abandoning the machine learning approach altogether.
• D. Analyze the recent data to understand the nature of the concept drift, retrain the model with a combination of historical and recent data, version the new model, and perform AIB testing against the existing model before fully deploying the new version. Log both model version predictions during AIB testing.
• E. Immediately replace the current UDF with a newly trained model using the latest data, ignoring model versioning and assuming the latest data will solve the drift issue.

A financial institution aims to detect fraudulent transactions using a Supervised Learning model deployed in Snowflake. They have a dataset with transaction details, including amount, timestamp, merchant category, and customer ID. The target variable is 'is_fraudulent' (0 or 1). They are considering different Supervised Learning algorithms. Which of the following algorithms would be MOST suitable for this fraud detection task, considering the need for interpretability, scalability, and the potential for imbalanced classes, and what specific strategies can be employed within Snowflake to handle the class imbalance?

• A. Support Vector Machine (SVM) with a radial basis function (RBF) kernel, as it can capture complex non-linear relationships without concern for interpretability.
• B. Naive Bayes, because it requires no hyperparameter tuning and works well on numerical data.
• C. Decision Tree or Random Forest, combined with techniques like oversampling the minority class (fraudulent transactions) within Snowflake using SQL or UDFs to balance the dataset before training. These models provide reasonable interpretability and can handle non-linear relationships effectively.
• D. Linear Regression, because it's computationally efficient and easy to understand, even though fraud detection is a classification problem.
• E. K-Nearest Neighbors (KNN), because it is simple to implement and doesn't require extensive training.

You are tasked with deploying a pre-trained sentiment analysis model hosted externally using AWS SageMaker. The model endpoint requires an API key for authentication, and you want to score customer reviews stored in a Snowflake table named 'CUSTOMER REVIEWS. Which of the following steps are necessary to securely and efficiently integrate this external model with Snowflake, assuming you have already created a Snowflake stage to store secrets?

• A. Create an external function in Snowflake that invokes the SageMaker endpoint, hardcoding the API key directly into the function definition for simplicity.
• B. Create an external function in Snowflake that retrieves the API key from a secure Snowflake secret object. Grant USAGE privilege on the secret to the service account associated with the external function.
• C. Create a secret object in Snowflake to store the API key. Grant appropriate privileges on the secret to the role that will execute the external function. Modify external function that references secure external stage.
• D. Store the API key in an environment variable within the AWS Lambda function (if using API Gateway) that serves as an intermediary between Snowflake and SageMaker. Snowflake calls the API Gateway endpoint which relays the request to the SageMaker endpoint, and no specific configuration is needed on snowflake.
• E. Use Snowflake's external functions to directly call the SageMaker endpoint from a SQL query, passing the customer review text as input. No separate secure external stage configuration is needed as long as Snowflake has internet access.

Your team has deployed a machine learning model to Snowflake for predicting customer churn. You need to implement a robust metadata tagging strategy to track model lineage, performance metrics, and usage. Which of the following approaches are the MOST effective for achieving this within Snowflake, ensuring seamless integration with model deployment pipelines and facilitating automated retraining triggers based on data drift?

• A. Leveraging a third-party metadata management tool that integrates with Snowflake and provides a centralized repository for model metadata, lineage tracking, and data governance. This tool should support automated tag propagation and data drift monitoring. Use Snowflake external functions to trigger alerts based on metadata changes.
• B. Using Snowflake's built-in tag functionality to tag tables, views, and stored procedures related to the model. Implementing custom Python scripts using Snowflake's Python API (Snowpark) to automatically apply tags during model deployment and retraining based on predefined rules and data quality checks.
• C. Utilizing Snowflake's INFORMATION SCHEMA views to extract metadata about tables, views, and stored procedures, and then writing custom SQL scripts to generate reports and track model lineage. Combine this with Snowflake's data masking policies to control access to sensitive metadata.
• D. Relying solely on manual documentation and spreadsheets to track model metadata, as automated solutions introduce unnecessary complexity and potential errors.
• E. Storing model metadata in a separate relational database (e.g., PostgreSQL) and using Snowflake external tables to access the metadata information. Implement custom stored procedures to synchronize metadata between Snowflake and the external database.