Snowflake SnowPro Advanced: Data Scientist Certification Sample Questions:

1. You've built a regression model in Snowflake using Snowpark Python to predict customer churn. After evaluating the model on a holdout dataset, you generate a residuals plot. The plot shows a distinct 'U' shape. Which of the following interpretations and subsequent actions are most appropriate?

A) The 'U' shape indicates homoscedasticity. No changes to the model are necessary.
B) The 'U' shape suggests the model is missing important non-linear relationships. Consider adding polynomial features or using a non-linear model like a Random Forest or Gradient Boosting Machine.
C) The 'U' shape indicates that the residuals are normally distributed. This is a positive sign and no changes are required.
D) The 'U' shape implies multicollinearity is present. Use techniques like Variance Inflation Factor (VIF) to identify and remove highly correlated features.
E) The 'U' shape suggests that the learning rate is too high. Reduce the learning rate of the model.

2. You have built and deployed a model to predict the likelihood of loan default using Snowpark and deployed as a Snowflake UDF. You are using a separate Snowflake table 'LOAN APPLICATIONS' as input, which contains current applicant data'. After several weeks in production, you observe that the model's accuracy has significantly dropped. The original training data was collected during a period of low interest rates and stable economic conditions. Which of the following strategies are the MOST effective for identifying potential causes of this performance degradation and determining if a model retrain is necessary, in the context of Snowflake?

A) Assume the model is no longer valid due to changing economic conditions and immediately retrain the model with the latest available data without further investigation.
B) Regularly sample data from the ' LOAN_APPLICATIONS table and manually compare it to the original training data. This provides a qualitative assessment of potential changes.
C) Re-run the original model training code with the 'LOAN_APPLICATIONS table as input and compare the resulting model coefficients to the coefficients of the deployed model. Significant differences indicate model decay.
D) Compare the distribution of input features in the 'LOAN_APPLICATIONS table to the distribution of the features in the original training dataset using Snowflake's statistical functions (e.g., APPROX_COUNT DISTINCT, &AVG', 'STDDEV'). Significant deviations indicate data drift.
E) Monitor the model's precision and recall using a dedicated monitoring dashboard built on top of the model's predictions and actual loan outcomes (once available). Create a Snowflake alert that triggers when either metric falls below a predefined threshold.

3. You are analyzing website clickstream data stored in Snowflake to identify user behavior patterns. The data includes user ID, timestamp, URL visited, and session ID. Which of the following unsupervised learning techniques, combined with appropriate data transformations in Snowflake SQL, would be most effective in discovering common navigation paths followed by users? (Choose two)

A) Sequence clustering using time-series analysis techniques (e.g., Hidden Markov Models), after transforming the data into a sequence of URLs for each session using Snowflake's LISTAGG function ordered by timestamp.
B) Principal Component Analysis (PCA) to reduce the dimensionality of the URL data, followed by hierarchical clustering. This will group similar URLs together.
C) DBSCAN clustering on the raw URL data, treating each URL as a separate dimension. This will identify URLs that are frequently visited by many users.
D) Association rule mining (e.g., Apriori) applied directly to the raw URL data to find frequent itemsets of URLs visited together within the same session. No SQL transformations are required.
E) K-Means clustering on features extracted from the URL data, such as the frequency of visiting specific domains or the number of pages visited per session. This requires feature engineering using SQL.

4. You are using Snowpark for Python to build a feature engineering pipeline for a machine learning model that predicts customer churn. The data is stored in a Snowflake table called 'CUSTOMER DATA' , and you want to create new features based on time-series data within the table. You need to calculate the 'Recency' feature (days since the last transaction) and 'Frequency' feature (number of transactions in the last 3 months). Considering performance and best practices, which Snowpark approach would you choose?

A) Write a stored procedure in SQL that calculates 'Recency' and 'Frequency' using SQL window functions, and then call this stored procedure from your Snowpark Python code.
B) Write custom Python code in a Snowpark UDF to retrieve each transaction for a customer and calculate recency and frequency directly in Python without pandas.
C) Use Snowpark DataFrame API to perform window functions within Snowflake to calculate 'Recency' and 'Frequency' directly, leveraging Snowflake's processing power without transferring data to the client.
D) Create a Python UDF using Pandas to calculate 'Recency' and 'Frequency'. Apply this UDF to the 'CUSTOMER DATA' table through Snowpark, processing the data row by row.
E) Fetch the entire 'CUSTOMER DATA table into a Pandas DataFrame using , then use Pandas' time-series functions to calculate 'Recency' and 'Frequency'. After feature engineering, load the Pandas DataFrame back into Snowflake.

5. You have a table 'PRODUCT SALES in Snowflake with columns: 'PRODUCT (INT), 'SALE_DATE (DATE), 'SALES_AMOUNT (FLOAT), and 'PROMOTION FLAG' (BOOLEAN). You need to perform the following data preparation steps using Snowpark SQLAPI:

A) Creating a feature that returns 1 if there is a PROMOTION_FLAG of True and SALES_AMOUNT > 1000, and zero otherwise
B) Handling missing 'SALES_AMOUNT values by imputing them with the average 'SALES_AMOUNT' for the same 'PRODUCT_ID during the previous month. If there's no data for the previous month, use the overall average for that
C) All of the above.
D) Converting 'SALE_DATE to a quarterly representation (e.g., '2023-QI').
E) Creating a new feature representing the percentage change in 'SALES_AMOUNT compared to the previous day for the same 'PRODUCT_ID. Handle the first day of each 'PRODUCT by setting 'SALES_GROWTH' to O.

Solutions: