Prepare and test your skills

Prepare and test your skills

Free sample question with answers and detailed explanation below. The Cloud Digital Leader practice bank has an additional 3,070 practice questions.Try 10 free realistic questions: instant score, no credit card, and no account required.
A retail organization is training a machine learning model to predict customer purchasing behavior. To accelerate the project timeline, the team imports raw historical transaction records directly into the training pipeline without preparation. The dataset contains the following issues:
When evaluating the trained model, the organization finds that its predictions are erratic and unreliable for guiding marketing decisions.
Which action should the team take to address the root cause of the model's poor performance?
Batch inference and online inference are the two primary methods for generating predictions from a deployed machine learning model. Batch inference processes a large volume of requests asynchronously and outputs results to a file or table, whereas online inference processes individual requests synchronously in real time via a REST API.
The method used to request predictions in production has no effect on how well the model learned during training. Whether inferences are requested asynchronously in bulk or synchronously one at a time, a model trained on missing values, inconsistent text, and noisy identifier columns will produce the exact same erratic and unreliable predictions.
A common misconception is that deployment infrastructure or API latency affects the statistical quality of machine learning inferences. In reality, model quality is determined by the integrity, relevance, and completeness of the training dataset. Deploying a flawed model to a real-time endpoint simply delivers unreliable predictions faster.
In classification models, each prediction is assigned a numeric confidence score that reflects the model's certainty that a predicted class is correct. The score threshold is the cutoff value used to convert that probability into a binary decision (such as "yes, the customer will purchase" or "no, they will not").
Adjusting the score threshold only changes how the model's output probabilities are filtered after training; it cannot repair a model that was trained on erroneous, incomplete, or noisy data. Furthermore, lowering the score threshold does not increase prediction accuracy or confidence—instead, it causes the model to accept lower-confidence predictions as positive, which significantly increases the risk of misclassification and false positives.
When a model behaves unexpectedly or produces erratic confidence scores during evaluation, debugging the model requires inspecting and improving the training data rather than manipulating evaluation sliders. A threshold adjustment is only meaningful when the model's underlying probability estimates are built on high-quality, representative data.
A test dataset (typically around 10% of the total dataset) is a completely separate slice of data held back from the model training process. Its primary purpose is to serve as an entirely new, unseen challenge once training is finished, providing an objective evaluation of how the machine learning model will perform on real-world data.
Combining the test set with the training set violates fundamental machine learning evaluation principles and does not fix inaccurate, incomplete, or noisy data. Feeding more uncleaned, erroneous records into the neural network simply reinforces the Garbage In, Garbage Out (GIGO) problem, causing the model to memorize flawed patterns and random noise rather than learning valid business logic.
Standard machine learning workflows divide data into three distinct subsets: the training set (to learn model weights), the validation set (to tune structural hyperparameters), and the test set (to evaluate final model performance). If a model performs poorly on the test set due to noisy or inconsistent features, teams must return to the data preparation phase to clean the dataset across all splits.
Data preparation and data cleaning are foundational steps in the machine learning lifecycle that ensure the input dataset is accurate, complete, and consistent before model training begins. This practice directly addresses the Garbage In, Garbage Out (GIGO) principle, which states that a machine learning model is only as reliable as the quality of the data used to train it. When a model is fed flawed, incomplete, or noisy inputs, it cannot identify true underlying patterns, resulting in compromised performance and diminished business value.

Machine learning algorithms cannot automatically distinguish between meaningful business patterns and data entry mistakes. Attempting to tune model evaluation settings or alter deployment architectures without fixing underlying data quality issues will still yield unreliable predictions. Cleaning the dataset directly eliminates the root cause of the poor model performance.
Answer them yourself and get your score the moment you finish. No account needed.
Full access includes
Keep the momentum going with these hand-picked practice scenarios
Last updated