DAA-C01 Question 106
Single answerClassificationA retail analytics team is building a churn prediction model in Snowflake using historical customer data stored in a feature table. The target column, CHURNED, contains values 0 and 1. The team wants to train and evaluate a model directly in Snowflake using SQL, and they want metrics that are appropriate for a binary classification problem. Which approach should they use?
- A
Use CREATE SNOWFLAKE.ML.CLASSIFICATION to train the model on the feature columns and CHURNED as the label, then evaluate it with classification metrics such as precision, recall, and ROC AUC.
- B
Use CREATE SNOWFLAKE.ML.FORECAST to predict CHURNED because churn is a future event, then evaluate it with MAPE and RMSE.
- C
Use CREATE SNOWFLAKE.ML.ANOMALY_DETECTION because customers who churn are rare, then evaluate the output with reconstruction error.
- D
Use CREATE SNOWFLAKE.ML.REGRESSION to predict CHURNED as a numeric value, then round predictions to 0 or 1 and evaluate with R-squared.
Show answer and explanation
Correct answer: A
Explanation
This scenario is a classic binary classification problem: the label column has two possible classes, 0 and 1. In Snowflake, the correct applied approach is to use the classification functionality in Snowflake ML for supervised learning on labeled data. For model evaluation, practitioners should use classification-oriented metrics such as precision, recall, F1, accuracy, and ROC AUC rather than regression or forecasting metrics. A common misconception is to choose forecasting because churn happens in the future, but forecasting is for time-series values indexed over time, not for categorical outcomes attached to individual training rows. Another misconception is to use anomaly detection because churn events may be less frequent; however, when labeled examples exist, supervised classification is the best-practice choice. This aligns with Snowflake ML documentation and general machine learning best practices for categorical target prediction.
- A. Correct.
Correct. CHURNED is a binary target, so this is a standard supervised classification use case. In Snowflake ML, classification methods are intended for categorical labels such as 0/1 churn outcomes. Metrics like precision, recall, F1, and ROC AUC are appropriate because they measure classification quality, including performance on potentially imbalanced classes.
- B. Incorrect.
Incorrect. Forecasting is intended for time-series prediction where the model learns temporal patterns over a time index. Although churn may happen in the future, the target here is a binary label per customer record, not a time-series forecasting problem. MAPE and RMSE are regression/forecast error metrics and are not the best fit for binary classification.
- C. Incorrect.
Incorrect. Anomaly detection is typically used when labeled outcomes are unavailable and the goal is to identify unusual records. Here, the team already has a labeled target column, CHURNED, which makes supervised classification the correct approach. Treating churn as an anomaly problem would ignore available labels and usually produce a less suitable model.
- D. Incorrect.
Incorrect. Regression predicts continuous numeric outcomes. While a practitioner might be tempted to model 0/1 values with regression and then threshold or round them, that is not the recommended approach when a true categorical label is available. R-squared is not an appropriate primary metric for binary classification performance.