Snowflake SnowPro Advanced: Data Scientist Certification : DSA-C03 Exam Questions

You have a structured dataset in Snowflake containing customer information and purchase history. You aim to build a multi-class classification model to predict customer churn, categorizing customers into 'Low Risk', 'Medium Risk', and 'High Risk' of churning. After training the model, you want to evaluate its performance. Which of the following metrics and evaluation techniques, when used together, provide the MOST comprehensive understanding of the model's performance across all churn risk categories, especially when dealing with potential class imbalance?

• A. Overall Accuracy, Precision, Recall, F I-Score for each class, and Confusion Matrix.
• B. Area Under the ROC Curve (AUC-ROC) for each class (one-vs-rest approach), Precision-Recall Curve for each class, and Cumulative Accuracy Profile (CAP) curve.
• C. Log Loss (Cross-Entropy Loss), Gini Coefficient, and Kolmogorov-Smirnov (KS) statistic.
• D. Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (Coefficient of Determination).
• E. Only Overall Accuracy and a confusion Matrix.

You are working with a large dataset in Snowflake and need to build a machine learning model using scikit-learn in Python. You want to leverage Snowflake's compute resources for feature engineering to speed up the process. Which of the following approaches correctly combines Snowflake's SQL capabilities with scikit-learn for feature engineering and model training, while minimizing data transfer between Snowflake and the Python environment?

• A. Use the Snowflake Python Connector to execute individual SQL queries for each feature engineering step. Load the resulting features step-by-step into a Pandas DataFrame and train the scikit-learn model.
• B. Create Snowflake User-Defined Functions (UDFs) in Python for complex feature engineering calculations. Call these UDFs within a SQL query to apply the feature engineering to the Snowflake data. Load the resulting features into a Pandas DataFrame and train the scikit-learn model.
• C. Write a complex SQL query in Snowmake to perform all feature engineering, then load the resulting features into a Pandas DataFrame and train the scikit-learn model.
• D. Use Snowflake external functions to invoke a remote service (e.g., AWS Lambda) for feature engineering. Pass data from Snowflake to the remote service, receive the engineered features back, and load them into a Pandas DataFrame for model training.
• E. Implement the feature engineering steps directly in Python using Pandas and scikit-learn, then load the raw data into a Pandas DataFrame and apply the transformations. Finally, train the scikit-learn model.

You are tasked with identifying fraudulent transactions from unstructured log data stored in Snowflake. The logs contain various fields, including timestamps, user IDs, and transaction details embedded within free-text descriptions. You plan to use a supervised learning approach, having labeled a subset of transactions as 'fraudulent' or 'not fraudulent.' Which of the following methods best describes the extraction and processing of this data for training a machine learning model within Snowflake?

• A. Treat the unstructured log description as a categorical feature and directly apply one-hot encoding within Snowflake, then train a classification model. Due to high dimensionality perform PCA for dimensionality reduction before training.
• B. Extract the entire log description field and train a word embedding model (e.g., Word2Vec) on the entire dataset. Average the word vectors for each transaction's log description to create a document vector. Train a classification model (e.g., Random Forest) on these document vectors within Snowflake.
• C. Export the entire log data to an external machine learning platform (e.g., AWS SageMaker) and perform feature extraction, NLP processing, and model training there. Import the trained model back into Snowflake as a UDF for prediction.
• D. Use regular expressions within a Snowflake UDF to extract relevant information (e.g., amount, item description) from the log descriptions. Convert extracted data into numerical features using one-hot encoding within the UDF. Then, train a model using the extracted numerical features directly within Snowflake using SQL extensions for machine learning.
• E. Use a combination of regular expressions and natural language processing (NLP) techniques within Snowflake UDFs to extract key features such as transaction amounts, product categories, and sentiment scores from the log descriptions. Then, combine these extracted features with other structured data (e.g., user demographics) and train a classification model using these features. The NLP steps include tokenization, stop word removal, and TF-IDF vectorization.

You are analyzing sensor data collected from industrial machines, which includes temperature readings. You need to identify machines with unusually high temperature variance compared to their peers. You have a table named 'sensor _ readings' with columns 'machine_id', 'timestamp', and 'temperature'. Which of the following SQL queries will help you identify machines with a temperature variance that is significantly higher than the average temperature variance across all machines? Assume 'significantly higher' means more than two standard deviations above the mean variance.

• A. Option C
• B. Option A
• C. Option B
• D. Option E
• E. Option D

A marketing analyst is building a propensity model to predict customer response to a new product launch. The dataset contains a 'City' column with a large number of unique city names. Applying one-hot encoding to this feature would result in a very high-dimensional dataset, potentially leading to the curse of dimensionality. To mitigate this, the analyst decides to combine Label Encoding followed by binarization techniques. Which of the following statements are TRUE regarding the benefits and challenges of this combined approach in Snowflake compared to simply label encoding?

• A. Binarizing a label encoded column using a simple threshold (e.g., creating a 'high_city_id' flag) addresses the curse of dimensionality by reducing the number of features to one, but it loses significant information about the individual cities.
• B. Binarization following label encoding may enhance model performance if a specific split based on a defined threshold is meaningful for the target variable (e.g., distinguishing between cities above/below a certain average income level related to marketing success).
• C. Label encoding introduces an arbitrary ordinal relationship between the cities, which may not be appropriate. Binarization alone cannot remove this artifact.
• D. While label encoding itself adds an ordinal relationship, applying binarization techniques like binary encoding (converting the label to binary representation and splitting into multiple columns) after label encoding will remove the arbitrary ordinal relationship.
• E. Label encoding followed by binarization will reduce the memory required to store the 'City' feature compared to one-hot encoding, and Snowflake's columnar storage optimizes storage for integer data types used in label encoding.