50 IBM A1000-080: Assessment: Data Science and AI Practice Questions: Question Bank 2025

A data scientist is asked to explain why a model’s performance drops when evaluated on new customer data compared to the training results. Which concept best describes this situation?

A team needs to build a model that predicts whether a transaction is fraudulent (yes/no) based on historical labeled data. Which type of machine learning problem is this?

A business stakeholder asks what the F1-score represents for a binary classifier. Which explanation is most accurate?

A developer wants to call an IBM Watson service from an application and needs a secure way to authenticate API requests. Which approach is recommended?

A dataset contains missing values in several numeric columns. The team plans to train a logistic regression model. Which preprocessing approach is generally appropriate?

A model for predicting equipment failures achieves 98% accuracy, but failures are rare (about 1% of records). Operations reports the model misses most failures. What is the best next step to assess model quality?

A team trains a deep neural network for image classification and observes training loss decreasing while validation loss starts increasing after several epochs. Which technique is most directly aimed at addressing this issue?

A data science team wants their model training and evaluation steps to be reproducible and auditable across environments. Which practice best supports this goal?

A team builds a sentiment classifier and sees unexpectedly high validation performance. Later, they discover the preprocessing step computed text features (e.g., vocabulary statistics) using the full dataset before splitting into train and validation sets. What is the primary issue, and what is the correct fix?

A company is deploying an AI model that influences loan approvals. They need to reduce the risk of discriminatory outcomes and provide transparency to auditors. Which approach is most appropriate?

A data analyst is asked to summarize the "typical" customer order value. The order values are highly skewed with a few very large purchases. Which metric is the best choice to represent a typical value?

A team is building a binary classifier to detect rare fraudulent transactions (1% fraud rate). They report 99% accuracy on a validation set. Which additional metric is most important to evaluate model usefulness for fraud detection?

A data scientist needs to split a labeled dataset into training and test sets while preserving the class distribution across splits. What is the recommended approach?

A model performs well during cross-validation but fails in production because the distribution of incoming data changed (e.g., new customer segment). What issue is most likely occurring?

A product team wants to group customers into segments without pre-labeled outcomes. Which approach is most appropriate?

A team trains a deep neural network and observes that training loss continues to decrease while validation loss starts increasing after several epochs. What is the most likely cause and best next step?

A binary classifier outputs probabilities. The business wants to reduce false negatives (missing positive cases), even if false positives increase. What should the team adjust?

A team is using IBM Watson Studio to build a model collaboratively. They need reproducible environments and a consistent way to share data preparation and training steps. Which artifact best supports this goal?

A model is trained to predict customer churn. During model development, the feature set includes "days until account closure" which is derived from a future event that occurs after the prediction time. Validation metrics look exceptionally high. What is the most likely problem?

A company wants an architecture for an AI solution where model training happens in a controlled environment, and the deployed model is invoked by applications through a stable interface with monitoring for drift and performance. Which design best matches this requirement?

A data scientist is asked to explain to a business stakeholder why a model's accuracy improved after adding a new feature. Which statement best describes what likely happened?

A team wants to quickly sanity-check whether their dataset has enough labeled examples per class before training a classifier. What is the most appropriate first step?

A project team wants a common framework to describe the main phases of a data science project from business understanding through deployment. Which methodology is most commonly used for this purpose?

A classifier shows 95% accuracy, but the business cares most about catching rare fraud cases. Which metric is typically most appropriate to prioritize for this use case?

A model performs very well on training data but noticeably worse on validation data. Which situation best describes what is happening, and what is a typical mitigation?

A team is building a model to predict house prices. The target variable is highly right-skewed with a few extreme outliers. Which approach is often a good first step to improve modeling stability?

A computer vision model is trained to classify products. During inference, the model performs poorly and seems sensitive to lighting changes, even though training accuracy was high. Which improvement is most directly targeted at this issue?

A team uses an 80/20 train-test split and performs feature scaling using the mean and standard deviation computed from the full dataset before splitting. Test performance looks unusually strong. What is the most likely issue?

A language model is being used to summarize internal customer emails. Some summaries include sensitive personal information that should be excluded. Which approach best addresses this requirement in a production AI workflow?

An organization wants to operationalize ML models and ensure reproducibility across teams. Which practice best supports this goal?

A data scientist is asked to clearly describe what an outlier is to a non-technical stakeholder. Which statement is the best definition?

A team wants to avoid data leakage when creating features for a churn model. Which approach is the best practice?

You trained a binary classifier and the business cares most about reducing false negatives (missing positive cases). Which metric should you prioritize?

An image classification model performs well on training data but poorly on new images. Training loss is low while validation loss is high and increasing. What is the most likely issue?

A team needs a confusion matrix for a multiclass model in Watson Studio. They already have predicted labels and true labels. What is the most direct way to generate it?

A model shows high accuracy but the positive class is rare and the model misses most positive cases. Which evaluation approach best addresses this?

A project requires traceability for model artifacts and reproducibility of experiments across team members. Which practice best supports this in an IBM-focused workflow?

A deep learning model for text classification is slow to train and overfits quickly. Which change is most likely to improve generalization without fundamentally changing the task?

A team deploys a model and notices performance degradation over time as customer behavior changes. Which action best addresses the root cause?

A stakeholder asks why a random train/test split may be inappropriate for time-ordered sales data. What is the best explanation?

A data scientist is asked to explain why the average customer satisfaction score increased last quarter. The dataset contains multiple regions, and one region’s score rose sharply due to a small number of customers. Which pitfall is most likely causing a misleading overall conclusion?

A team trains a classifier on a dataset where 98% of the records are non-fraud and 2% are fraud. They report 98% accuracy and claim the model is excellent. Which evaluation metric is most appropriate to challenge this claim for the minority (fraud) class?

In IBM Watson Studio, a user wants to ensure collaborators can reproduce a notebook run with the same library dependencies. Which practice best supports reproducibility?

A model performs well in training but poorly in production after a marketing campaign changes customer behavior. The feature distributions have shifted, while the label definition remains the same. What is the most likely issue?

A retailer wants to segment customers into groups based on purchasing behavior for targeted campaigns. They do not have labeled outcomes for training. Which approach is most appropriate?

A deep learning image classifier shows strong performance on the training set, but validation accuracy is much lower. The team has limited data. Which technique is most likely to improve generalization without collecting new labeled images?

A team is building a churn model and wants to prevent leaking future information. They currently include a feature called "number_of_support_tickets_next_30_days". What is the best action?

A team needs a repeatable pipeline to train, evaluate, and store an ML model artifact with traceability (data version, code version, metrics) and enable future deployment. Which best-practice approach aligns with IBM tools and MLOps principles?

A team evaluates a binary classifier using ROC-AUC and gets 0.92, but the model still performs poorly for the business because missing positive cases is very costly. Which adjustment is most appropriate to address the business objective while keeping the same model?

An NLP team fine-tunes a transformer model for sentiment analysis. Training loss decreases steadily, but validation loss starts increasing after a few epochs. Which combination of actions is MOST appropriate to mitigate the issue?