ML-ASSOC sample questions

Question 1

A data scientist is cleaning a dataset containing employee salary information for a large corporation. They observe that the 'salary' column has a number of extreme outliers, including several C-level executive salaries that are orders of magnitude higher than the rest of the employees. For a feature engineering task, they need to impute a few missing salary values. Which imputation method should they prefer and why?

Accepted Answer

Median, because it is robust to extreme outliers.

Answer

Mean, because it provides the most accurate measure of central tendency.

Answer

Mode, because it represents the most common salary.

Answer

A constant value of zero, to clearly indicate that the data was missing.

Question 2

A leading e-commerce company wants to deploy a new product recommendation system. The system has several complex requirements:

1. **Real-time Personalization**: The model must provide recommendations within 200ms of a user's action (e.g., viewing a product).
2. **Dynamic User Profiles**: User feature vectors, which are used for inference, must be updated in near real-time based on their clickstream data.
3. **A/B Testing**: The MLOps team must be able to deploy a new 'challenger' recommendation algorithm and route 10% of live traffic to it for evaluation against the current 'champion' model.
4. **Scalability**: The system must handle traffic spikes during holiday seasons, scaling automatically without manual intervention.

Which Databricks deployment architecture best fulfills all these requirements?
```mermaid
graph TD
subgraph User Interaction
WebApp[Web Application] --> API_GW[API Gateway]
end
subgraph Real-time Processing
Clickstream[Kafka: User Events] --> DLT[Delta Live Tables: User Profile Update]
DLT --> OnlineStore[Online Feature Store]
end
subgraph Model Serving
API_GW --> Endpoint[Model Serving Endpoint]
Endpoint -- 90% --> Champion[Champion Model]
Endpoint -- 10% --> Challenger[Challenger Model]
Champion --> OnlineStore
Challenger --> OnlineStore
end
subgraph Batch Processing
ProductCatalog[Batch: Product Catalog] --> OfflineStore[Offline Feature Store]
end
```

Accepted Answer

Deploy both champion and challenger models to a single serverless Model Serving endpoint, configure traffic splitting, and use an Online Feature Store for low-latency feature lookups during inference.

Answer

Use a daily batch inference job that pre-computes recommendations for all users and stores them in a key-value store for fast retrieval.

Answer

Deploy the model as a Spark UDF within a Delta Live Tables pipeline to process user requests in micro-batches.

Answer

Create two separate Model Serving endpoints, one for the champion and one for the challenger, and manage traffic splitting using an external application load balancer.

Question 3

A machine learning engineer is using the `FeatureEngineeringClient` to create a new feature table in Unity Catalog. The table will store user features, and it's critical that each user is uniquely identified and that features can be looked up efficiently for online serving. Which parameter in the `fe.create_table` method is used to specify the unique identifier column(s) for the entities in the table?

Accepted Answer

primary_keys

Answer

partition_cols

Answer

index_cols

Answer

timeseries_columns

Question 4

A data scientist is using Hyperopt to perform Bayesian hyperparameter optimization for a machine learning model. They need to supply the correct search algorithm to the `algo` parameter of the `fmin` function. Which of the following options implements a Bayesian approach, specifically Tree-structured Parzen Estimator?

Accepted Answer

tpe.suggest

Answer

rand.suggest

Answer

grid.suggest

Answer

anneal.suggest

Question 5

An MLOps team is managing a critical fraud detection model registered in Unity Catalog. The current production model is aliased as 'Champion'. A new version has been validated and is ready to be promoted. To minimize risk, the team needs a strategy that allows for an immediate rollback to the previous version if the new model underperforms. Which steps should they perform? (Select TWO)

Accepted Answer

Set the 'Champion' alias on the new model version.

Accepted Answer

Set a 'Previous_Champion' alias on the old model version.

Answer

Delete the 'Champion' alias from the old model version.

Answer

Delete the old model version from the registry to avoid confusion.

Answer

Retrain the old model version with the newest data.

Question 6

During exploratory data analysis for a demand forecasting model, a data scientist observes that a key feature, `user_daily_logins`, has a strong positive skew, with most users logging in 1-2 times a day but a small number of power users logging in over 50 times. This skew could negatively impact the performance of a linear regression model. Which feature transformation is most appropriate to apply to this feature?

Accepted Answer

Logarithmic transformation

Answer

Standardization (Z-score scaling)

Answer

One-hot encoding

Answer

Normalization (Min-Max scaling)

Question 7

A machine learning team is using `GridSearchCV` from scikit-learn to tune a Gradient Boosting model. The parameter grid is defined as follows:
`param_grid = {'n_estimators': [100, 200], 'learning_rate': [0.01, 0.1, 0.2], 'max_depth': [3, 5, 7]}`
They are using 5-fold cross-validation (`cv=5`). How many individual models will be trained during this entire hyperparameter tuning process?

Accepted Answer

90

Answer

18

Answer

23

Answer

45

Question 8

A logistics company needs to deploy a model for real-time anomaly detection in its package delivery event stream. The pipeline must process a high volume of events with fluctuating loads and requires a solution that simplifies infrastructure management and automatically handles cluster scaling. What is the primary advantage of using Delta Live Tables (DLT) for this streaming inference task compared to a manually configured Structured Streaming job?

Accepted Answer

DLT offers declarative pipeline development and managed autoscaling of infrastructure.

Answer

DLT provides lower latency for individual event processing.

Answer

DLT allows for model inference using Python, which Structured Streaming does not.

Answer

DLT can only be used for batch inference, not streaming.

Question 9

A small data science team is starting a new project to predict customer lifetime value. They have a well-prepared dataset but limited time for initial model exploration. What is the key benefit of using Databricks AutoML at the beginning of this project?

Accepted Answer

It quickly generates a strong baseline model and editable notebooks, accelerating the development cycle.

Answer

It guarantees the creation of a production-ready model with optimal performance.

Answer

It is the only tool that can handle datasets larger than 100 GB.

Answer

It automatically deploys the best model to a real-time serving endpoint.

Question 10

A fraud detection model is being developed for a financial institution. The dataset is highly imbalanced, with fraudulent transactions representing only 0.1% of the data. The primary business objective is to minimize financial losses, meaning the cost of a false negative (missing a fraudulent transaction) is thousands of times higher than the cost of a false positive (flagging a legitimate transaction). Which technique for handling data imbalance is most directly aligned with this business objective?

Accepted Answer

Implementing cost-sensitive learning by assigning a much higher misclassification weight to the minority class.

Answer

Randomly undersampling the majority class to create a balanced dataset.

Answer

Using SMOTE (Synthetic Minority Over-sampling Technique) to generate new minority class samples.

Answer

Adjusting the classification threshold of the model during post-processing to increase recall.

Free ML-ASSOC sample questions

A small data science team is starting a new project to predict customer lifetime value. They have a well-prepared dataset but limited time for initial model exploration. What is the key benefit of using Databricks AutoML at the beginning of this project?

10 more free samples are waiting