AI Associate Data for AI Practice Test 3 — 15 Questions

Question 1

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.. 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 2

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.. 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 3

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.. The error indicates the query references fields that are not present in the policy's allowed field list. In data-access policy systems (such as Snowflake masking policies, AWS Lake Formation, or similar), a query is rejected when it requests columns the policy does not expose, even if the user is otherwise authorized. The most likely cause is that the requested fields are not included in the policy definition.

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 4

An organization is preparing data for Einstein Next Best Action. They have multiple action types (discounts, product suggestions, content). Which data model approach best ensures accurate recommendations?

Accepted Answer

Use a single model that includes all action types in the training data.. Einstein Next Best Action is designed to learn from all action types simultaneously within a single model. By including all action types (discounts, product suggestions, content) in the training data, the model can capture cross-action patterns and relative effectiveness, leading to more accurate and contextually relevant recommendations. A unified model avoids fragmentation and ensures consistent scoring across actions.

Answer

Train a separate model per customer segment and then merge.

Answer

Create a separate model for each action type and combine results manually.

Answer

Build an ensemble of models and average their outputs.

Question 5

A data scientist is preparing data for Einstein Discovery. The dataset has 10,000 records with 5 predictors and one outcome. The outcome is binary (1/0). What is the minimum number of positive outcomes typically required for a reliable model?

Accepted Answer

250. For binary classification with 10,000 records and 5 predictors, a common rule of thumb in predictive modeling (including Einstein Discovery) is to have at least 10 events per predictor variable (EPV). With 5 predictors, you need at least 50 positive outcomes, but to ensure model stability and reliable training, a minimum of 250 positive outcomes (5% of 10,000) is typically required. This aligns with best practices for avoiding overfitting and achieving adequate statistical power.

Answer

500

Answer

100

Answer

50

Question 6

A company uses Einstein Forecasting for revenue prediction. The historical data shows seasonal spikes every quarter. The model consistently underestimates peak periods. What is the best data preparation step to improve accuracy?

Accepted Answer

Add a 'quarter' index field (1-4) to the dataset.. The model underestimates seasonal peaks because it lacks an explicit signal that identifies which quarter each data point belongs to. Adding a 'quarter' index field (1-4) gives the forecasting algorithm a categorical feature it can use to learn quarter-specific patterns, directly addressing the seasonal spikes. This is a standard feature engineering step for seasonal time series, enabling the model to adjust its predictions for known cyclical behavior.

Answer

Increase the forecast horizon to 12 months.

Answer

Remove the spike data points as outliers.

Answer

Use only the last 6 months of data to reduce noise.

Question 7

Which TWO data preparation steps are required before using Einstein Discovery for sales forecasting? (Choose 2)

Accepted Answer

Include a date or timestamp field for time series analysis. Option C is correct because Einstein Discovery's time series forecasting requires a date or timestamp field to establish the temporal ordering and intervals of the historical data it learns from. Option D is correct because Einstein Discovery cannot build reliable models when predictor variables contain missing values, so gaps must be handled (e.g., imputed or removed) before training. Option A is not required, since Einstein Discovery automatically handles categorical/text variables internally and does not force one-hot encoding. Option B is not a mandatory step, as duplicate removal is a general data-quality practice rather than a specific prerequisite for forecasting. Option E is not required, because Einstein Discovery does not require numeric predictors to be normalized to a 0-1 scale.

Answer

Convert all text fields to numeric using one-hot encoding

Answer

Remove duplicate records

Answer

Normalize numeric fields to a 0-1 scale

Question 8

Which THREE actions are recommended when preparing data for Einstein Next Best Action? (Choose 3)

Accepted Answer

Provide data on which actions were offered and whether they were accepted. Option A is correct because Einstein Next Best Action learns from the outcome of previously presented recommendations, so capturing which actions were offered and whether the customer accepted them provides the labeled training data (positive examples) the model needs to rank future recommendations. Option C is correct because recording rejections or actions not taken supplies negative examples, which are essential for the model to distinguish between effective and ineffective recommendations rather than only seeing successful outcomes. Option E is correct because recommendation relevance degrades as customer behavior and business context change, so retraining the model weekly with fresh interaction data keeps the predictions aligned with current patterns. Option B does not belong because there is no requirement of at least 10 action types per strategy; the number of actions should reflect the business use case, not a fixed minimum. Option D does not belong because restricting training data to only the last 30 days is not a recommended practice and could discard valuable historical signal; the guidance is to use sufficient, representative historical data rather than an arbitrary 30-day window.

Answer

Include at least 10 different action types per strategy

Answer

Use only historical data from the last 30 days

Question 9

Refer to the exhibit. A data scientist sees this error when training an Einstein Discovery model for customer churn prediction. What is the most likely reason for the error?

Accepted Answer

The positive examples (180) are insufficient for the number of fields (8).. Einstein Discovery requires a sufficient ratio of positive examples to predictor fields to build a statistically valid model. With only 180 positive examples across 8 fields, the model lacks enough signal to learn meaningful patterns, triggering the error. The platform enforces a minimum number of positive outcomes relative to field count to avoid overfitting and unreliable predictions.

Answer

The field count (8) exceeds the maximum of 5 allowed fields.

Answer

The model name contains a version number, which is not allowed.

Answer

The dataset has too few records (3200) for 8 fields.

Question 10

Refer to the exhibit. A developer runs this SOQL query to prepare data for Einstein Lead Scoring. The query returns an error. What is the most likely issue?

Accepted Answer

The alias 'TotalAmount' is not allowed in the HAVING clause.. In SOQL, aliases defined in the SELECT clause cannot be referenced in the HAVING clause. The HAVING clause must use the original aggregate expression, such as SUM(Amount), not the alias 'TotalAmount'. This is a common SOQL restriction that causes errors when developers try to use aliases for readability.

Answer

The query misses a GROUP BY clause.

Answer

The SUM(Amount) cannot be used in the HAVING clause.

Answer

The WHERE clause condition is invalid.

Question 11

Refer to the exhibit. A dataflow is set up to prepare data for a prediction model. The model is expected to predict close probability for all open opportunities. What is wrong with this dataflow?

Accepted Answer

The filter on StageName is too restrictive; it excludes non-won opportunities needed for training.. The filter excludes all opportunities that are not 'Closed Won'. The model should be trained on both won and lost opportunities to predict close probability. The filter should be removed or include all stages.

Answer

The output target should be a dataset, not a model.

Answer

The source should be Lead, not Opportunity.

Answer

The dataflow is missing a transform node to remove null values.

Question 12

A company wants to train an AI model to predict customer churn using historical data that contains many missing values. What is the best practice for handling missing data?

Accepted Answer

Impute missing values using mean or median.. Imputing missing values using mean or median is a standard practice that preserves the dataset size and statistical properties, allowing the AI model to learn from all available features without introducing bias from data removal. This approach is particularly effective for numerical features in customer churn prediction, where missing values are often random and imputation maintains the distribution for algorithms like logistic regression or gradient boosting.

Answer

Use only features without missing values.

Answer

Ignore missing values as they do not affect AI training.

Answer

Remove all records with missing values.

Question 13

What is the most likely cause of the error?

Accepted Answer

Data quality threshold violation. The most likely cause of the error is a data quality threshold violation. In AI systems, data quality thresholds are often set to ensure that input data meets certain criteria (e.g., completeness, accuracy) before processing. If the data falls below these thresholds, the system may raise an error to prevent unreliable predictions. This is a common safeguard in production AI pipelines.

Answer

Authentication failure

Answer

Data schema mismatch

Answer

Network timeout

Question 14

What is being performed in this command?

Accepted Answer

Batch prediction. The command shown is performing batch prediction (option B), which means applying an already-trained model to a large set of input records at once and writing the results to a destination such as a file or table, rather than training or transforming features. Batch prediction is distinct from model training (option C), which would involve fitting model parameters on labeled data, and from feature engineering (option A), which would create or transform input variables. It also differs from data validation (option D), which checks data quality, schema, or statistics rather than generating inference outputs. Therefore, the only fitting description of the operation is batch prediction.

Answer

Feature engineering

Answer

Model training

Answer

Data validation

Question 15

A company plans to train an AI model using data from Salesforce CRM and an external marketing automation platform. What is the first step to unify these data sources in Data Cloud?

Accepted Answer

Define a Data Model that maps fields from both sources to a unified customer object. Creating a data model that maps fields from both sources to a common object ensures consistency. Option B is wrong because data streams come after the model. Option C is wrong because Calculated Insights are for aggregations. Option D is wrong because data transformations are applied later.

Answer

Create two separate Data Streams to bring data in

Answer

Build a Calculated Insight to merge the data

Answer

Set up a Data Transformation to blend the sources