ML-ASSOC Machine Learning Associate Practice Questions

This is the core distinction. An Estimator (e.g., LogisticRegression, StringIndexer) has a .fit() method that takes a DataFrame and learns from it to produce a model, which is a Transformer. A Transformer (e.g., a trained model, StandardScalerModel, OneHotEncoder) has a .transform() method that takes a DataFrame and returns a new DataFrame with a column appended or changed.

This statement is true. Databricks Model Serving relies on the pyfunc model flavor as a standardized interface for loading models and serving predictions. This allows the serving infrastructure to handle any model type (scikit-learn, TensorFlow, custom Python models, etc.) in a generic way, as long as it's wrapped in the pyfunc flavor.

mlflow.log_artifact() is the correct function for logging arbitrary files, such as images, text files, or model plots, to an MLflow run. log_metric is for numerical values, log_param is for key-value parameters, and log_model is for saving the model itself in a specific format.

Because the model must be trained k times (once for each fold), the overall training time is roughly k times longer than a single train-test split.

For time-series data, random shuffling of data into folds (as done in standard k-fold) violates the temporal order. This means the model could be trained on future data to predict the past, leading to an overly optimistic performance estimate. Specialized techniques like TimeSeriesSplit are required.

Linear models, such as Linear and Logistic Regression, are sensitive to multicollinearity. When a categorical feature is one-hot encoded, the resulting binary columns are perfectly correlated (e.g., if a value is not in categories A, B, or C, it must be in D). This perfect correlation, known as the dummy variable trap, can make the model's coefficient estimates unstable and difficult to interpret. Dropping one of the encoded columns is a common way to resolve this.

Creating Feature Store tables within Unity Catalog at the account level is the correct approach. Unity Catalog provides a centralized governance and access control layer that spans across all workspaces within an account. This allows the central governance team to manage permissions on a single set of feature tables, which can then be securely accessed by authorized users and services from any workspace, fulfilling the core requirements of centralization and secure sharing.

This statement is true. Increasing model complexity allows the model to learn more intricate patterns from the training data, which reduces its bias (the error from erroneous assumptions in the learning algorithm). However, a more complex model is also more likely to learn the noise in the training data, making it more sensitive to variations in the data. This increased sensitivity is known as higher variance, which can lead to overfitting and poor generalization to new, unseen data.

The median is the correct choice because it is a robust measure of central tendency, meaning it is not significantly affected by extreme outliers. The mean, on the other hand, is sensitive to outliers; the very high executive salaries would pull the mean upwards, making it an unrepresentative value for the typical employee. Using the median ensures the imputed value is closer to the center of the majority of the data points.

This architecture correctly addresses all requirements. A serverless Model Serving endpoint provides low-latency (<200ms) inference and automatic scaling. Deploying multiple models to a single endpoint with traffic splitting directly enables A/B testing. An Online Feature Store, updated by a streaming process like DLT, provides the dynamic, low-latency feature vectors needed for real-time personalization.

The primary_keys parameter is used to specify the column or columns that uniquely identify each row or entity in the feature table. This is crucial for the Feature Store as it uses these keys for joining features during training set creation and for efficient lookups in online stores.