AI Associate Data for AI Practice Test 6 — 15 Questions

Question 1

A data scientist notices that the model accuracy drops significantly after retraining with new data. Upon inspection, they find that many records have missing values for a key feature. Which data quality improvement should be prioritized first?

Accepted Answer

Implement imputation for missing feature values.. The core issue is that missing values in a key feature introduce noise and bias, directly degrading model performance. Imputation (option A) is the most direct and impactful first step because it preserves the dataset size and feature set, allowing the model to learn from complete patterns. Without addressing missing data first, other quality improvements like normalization or feature reduction would be applied to corrupted data, failing to resolve the root cause.

Answer

Normalize the feature range.

Answer

Reduce the number of features.

Answer

Remove duplicate records.

Question 2

A company is building a text classification model for customer support tickets. They have a dataset of 10,000 tickets. The team decides to use active learning for labeling. Which approach best aligns with active learning principles?

Accepted Answer

Train a preliminary model and prioritize labeling tickets with low prediction confidence.. Active learning iteratively selects the most informative unlabeled data points for labeling, typically those with low prediction confidence from a preliminary model. This minimizes labeling effort while maximizing model performance, which is the core principle of active learning.

Answer

Randomly select 2,000 tickets and label them manually.

Answer

Use a pre-trained model to label all tickets automatically.

Answer

Have subject matter experts label all 10,000 tickets.

Question 3

For an AI project, data must be stored in a way that supports both training and real-time inference. Which storage solution meets this requirement?

Accepted Answer

Data lake (e.g., Amazon S3 or Azure Data Lake). A data lake (e.g., Amazon S3 or Azure Data Lake) is the correct choice because it can store vast amounts of raw, unstructured, and structured data in its native format, making it ideal for training AI models on diverse datasets. At the same time, data lakes support real-time inference by enabling direct access to data via APIs or streaming services (e.g., AWS Lambda or Azure Functions) without the latency of transforming data into a schema-on-write structure. This dual capability—handling both batch processing for training and low-latency reads for inference—is a key requirement that other storage solutions cannot fulfill as effectively.

Answer

Data warehouse (e.g., Snowflake)

Answer

Relational database (e.g., PostgreSQL)

Answer

In-memory cache (e.g., Redis)

Question 4

A data engineer needs to create a feature that represents the average purchase amount per customer over the last 30 days. The transactional data is timestamped. Which feature engineering technique is most appropriate?

Accepted Answer

Rolling average of purchase amounts over a 30-day window. Option B is correct because a rolling average over a 30-day window directly computes the average purchase amount per customer for only the most recent 30 days of transactions, which matches the requirement of a time-sensitive feature. This technique uses a sliding window function (e.g., AVG() with a ROWS or RANGE frame in SQL, or rolling().mean() in pandas) that respects the timestamp order, ensuring only relevant data contributes to the feature.

Answer

Sum of all purchase amounts per customer

Answer

Count of purchases per customer

Answer

Minimum purchase amount per customer

Question 5

A healthcare AI model uses patient data. The legal team requires that all data used for training be de-identified according to HIPAA Safe Harbor method. Which data handling process satisfies this?

Accepted Answer

Remove all 18 HIPAA identifiers from each record.. Option A is correct because the HIPAA Safe Harbor method specifically requires the removal of all 18 identifiers listed in the HIPAA Privacy Rule from each patient record. This includes direct identifiers like names, addresses, and Social Security numbers, as well as indirect identifiers such as dates and geographic subdivisions. By removing these 18 identifiers, the data is considered de-identified and no longer subject to HIPAA restrictions, allowing it to be used for AI training.

Answer

Generate synthetic data that mimics patient records.

Answer

Remove patient names and replace with IDs.

Answer

Anonymize data by aggregating into groups of 10 or more.

Question 6

A team is building a pipeline to train a model daily. The source data arrives in CSV files but needs to be converted to Parquet for efficiency. Which pipeline step should perform this conversion?

Accepted Answer

Data ingestion step. Option D is correct because the data ingestion step is responsible for bringing raw data into the pipeline, including format conversions like CSV to Parquet. Converting to Parquet at ingestion improves storage efficiency and query performance for downstream processing, as Parquet uses columnar storage and compression.

Answer

Feature engineering step

Answer

Model deployment step

Answer

Data validation step

Question 7

During data transformation, a data scientist applies one-hot encoding to a categorical feature with 50 unique values. The resulting dataset has 50 new columns. What is a potential drawback of this transformation?

Accepted Answer

High cardinality leading to sparse data and overfitting. One-hot encoding a categorical feature with 50 unique values creates 50 binary columns, each representing one category. This high cardinality leads to a very sparse matrix (most entries are 0), which can cause the model to overfit by learning noise from rare categories, especially when the dataset is not large enough to support such dimensionality.

Answer

Reduction in training time

Answer

Increased interpretability of the model

Answer

Loss of ordinal information in categories

Question 8

An organization uses Salesforce Data Cloud to unify customer data from multiple sources. They want to ensure that data lineage is tracked for AI models. Which practice supports data lineage?

Accepted Answer

Maintain metadata that records source, transformations, and dependencies.. Maintaining metadata that records source, transformations, and dependencies is the correct practice because data lineage for AI models requires a complete audit trail of where data originated, how it was transformed, and its dependencies. In Salesforce Data Cloud, this metadata is captured through the Data Catalog and Data Lineage feature, which tracks the flow of data from source objects through calculated insights and segments to AI model inputs, ensuring transparency and reproducibility.

Answer

Use data partitioning to improve query performance.

Answer

Implement role-based access control on datasets.

Answer

Regularly run data profiling to check completeness.

Question 9

A machine learning team is preparing a dataset for a supervised learning task. They have 100,000 labeled samples. Which data preparation step is essential before splitting into train/test sets?

Accepted Answer

Shuffle the dataset randomly.. Option C is correct because shuffling the dataset randomly before splitting into train/test sets ensures that the data distribution is similar across both subsets. Without shuffling, the split might inadvertently separate ordered or grouped data (e.g., time-series or batches), leading to biased model evaluation. This step is essential for supervised learning to prevent data leakage and ensure the test set is representative of the overall population.

Answer

Normalize all features to the same scale.

Answer

Remove all outliers from the dataset.

Answer

Visualize the data distribution for each feature.

Question 10

Which TWO of the following are common dimensions of data quality that must be addressed for AI training?

Accepted Answer

Accuracy of values. Accuracy of values (Option B) is a fundamental dimension of data quality because AI models learn patterns from training data; if the data contains incorrect values, the model will learn and propagate those errors, leading to unreliable predictions. For example, in a dataset of customer ages, a single erroneous entry of '200' can skew the model's understanding of age distributions, directly impacting model performance.

Answer

Storage efficiency

Answer

Encryption strength

Answer

Consistency with external benchmarks

Question 11

Which TWO considerations are critical when planning data labeling for a computer vision project in a regulated industry?

Accepted Answer

Mitigating labeler bias to ensure fairness. Option B is correct because labeler bias can introduce systematic errors into the training data, leading to models that perform unfairly or inaccurately across different demographic groups. In regulated industries, such bias can violate anti-discrimination laws and regulatory standards, making its mitigation a critical planning consideration.

Answer

Data storage location for label files

Answer

Labeling timeline and budget constraints

Answer

Choosing between bounding boxes and segmentation masks

Question 12

Which THREE types of data sources are commonly integrated into Salesforce Data Cloud for AI use cases?

Accepted Answer

Third-party demographic data. Option A is correct because Salesforce Data Cloud can ingest third-party demographic data from external sources (e.g., data enrichment providers) to enrich customer profiles. This data, when combined with first-party data, enables AI models to generate more accurate predictions and segmentations. Data Cloud’s Data Streams and Data Lake objects support structured ingestion of such external datasets.

Answer

Model training logs

Answer

Data transformation scripts

Question 13

Refer to the exhibit. A data scientist tries to query the dataset but receives an error. Which of the following is the most likely cause?

Accepted Answer

The requested fields are not included in the policy.. Option A is correct because 'data_scientist' is in allowedUsers, so they are allowed. Option B is not in policy, C the policy filters amounts >0 and <10000, so 5000 is included, D the fields are in the policy, so they should be accessible.

Answer

The condition filters out records with amount=5000.

Answer

The data scientist is not listed in the allowedUsers array.

Answer

The policy format is invalid JSON.

Question 14

Refer to the exhibit. A data pipeline fails during the DataTransformation stage. What is the most likely root cause?

Accepted Answer

A transformation step references the 'age' column, but it is not present in the input data.. Option C is correct because the error occurs during the DataTransformation stage, which processes data after it has been successfully ingested. If a transformation step references the 'age' column but that column is missing from the input data, the pipeline will fail with a column-not-found error. This is a common schema mismatch issue in data pipelines, distinct from connectivity or data quality problems.

Answer

The pipeline has a network connectivity issue.

Answer

The data type for 'income' is incorrect.

Answer

The 'age' column contains null values.

Question 15

Refer to the exhibit. A data transformation configuration is shown. Which of the following describes the outcome of applying this transformation?

Accepted Answer

'color' is one-hot encoded into multiple binary columns; 'price' and 'weight' are standardized to have mean 0 and variance 1.. Option C is correct because the transformation configuration applies a one-hot encoder to the 'color' categorical column, creating multiple binary columns, and applies a standard scaler to the 'price' and 'weight' numerical columns, centering them to mean 0 and scaling to unit variance. This is a common preprocessing pipeline that handles mixed data types appropriately.

Answer

The transformation is invalid because one-hot encoding cannot be combined with scaling.

Answer

Only 'color' is transformed; 'price' and 'weight' are unchanged.

Answer

'color' is scaled to [0,1] and 'price', 'weight' are one-hot encoded.