Hard Difficulty Questions

Encode them as one-hot vectors.

Why wrong: One-hot encoding with high cardinality creates too many features.

Exclude them from the model.

Why wrong: Excluding may lose valuable information.

Use the raw values without transformation.

Why wrong: Raw values with many levels can cause overfitting.

Remove zip code from the feature set and retrain.

Why wrong: Other features may still correlate with race.

Keep zip code but add a fairness penalty to the loss function.

Why wrong: Retaining proxy may still cause disparate impact.

Increase transparency by publishing the model's decision criteria.

Why wrong: Transparency does not correct bias.

Accuracy

Why wrong: The system may be accurate in predicting clicks.

Privacy

Why wrong: Privacy may not be violated if data is properly handled.

Transparency

Why wrong: Transparency alone does not prevent harm.

Remove race and age features entirely to ensure fairness.

Why wrong: Age may be clinically relevant; removal could reduce accuracy unfairly.

Replace age with an age group bucket to reduce granularity.

Why wrong: This does not address bias from race; still problematic.

Use the model as is because predictions are accurate.

Why wrong: Accuracy does not justify ethical or legal risk.

The data stream object is not a standard Salesforce object.

Why wrong: Custom objects can be used as data streams.

The data stream is not activated for identity resolution.

Why wrong: Activation is for using the data, not for unification.

The data source is not from Salesforce, so it cannot be unified.

Why wrong: Data Cloud supports multiple external sources.

Event data from the website tracked via Google Analytics.

Why wrong: Event data must be brought into Salesforce to be used.

Static profile fields like customer age and location.

Why wrong: These do not reflect recent behavior.

Historical data from a data warehouse updated daily.

Why wrong: Daily updates may not be timely enough for real-time personalization.

Apply Principal Component Analysis (PCA) for dimensionality reduction

Why wrong: PCA does not handle class imbalance.

Create interaction features between existing variables

Why wrong: Interaction features do not fix imbalance.

Use accuracy as the evaluation metric

Why wrong: Accuracy is misleading for imbalanced data.

Einstein features are not enabled in the org

Why wrong: Einstein is enabled by default for Professional and Enterprise editions with Sales Cloud.

Users do not have the 'View Scores' permission

Why wrong: Permission affects visibility, not score generation.

The accounts lack custom fields for scoring

Why wrong: Scoring uses standard and custom fields; lack of custom fields doesn't prevent scoring.

Create multiple prediction models and average their scores

Why wrong: Einstein Prediction Builder uses a single model; averaging is not supported.

Add more custom fields as predictor fields

Why wrong: More fields can help but only if relevant; quality matters more than quantity.

Increase the frequency of model retraining

Why wrong: Retraining frequency is automatic based on data changes.

Update account records via process builder

Why wrong: No impact on resolution time.

Configure a flow to send churn alerts

Why wrong: Alerts don't reduce churn.

Create a custom report on churn

Why wrong: Only visualizes existing data, not prescriptive.

Enable stemming and synonyms.

Why wrong: Stemming and synonyms improve recall, not necessarily ranking.

Increase the number of indexed fields.

Why wrong: More fields may add noise.

Decrease the index refresh frequency.

Why wrong: Refresh frequency affects freshness, not ranking.

Include race and gender as predictors to allow the model to adjust for them.

Why wrong: Including them can introduce or amplify bias.

Rely on the model's built-in fairness constraints.

Why wrong: Einstein Prediction Builder does not include fairness constraints.

Use a deep learning algorithm to automatically handle bias correction.

Why wrong: Deep learning does not automatically correct bias; careful data preparation is needed.

A fine-tuned large language model

Why wrong: Fine-tuning updates model weights but doesn't supply real-time data; it's a static improvement.

A prompt template in Prompt Builder

Why wrong: Prompt Builder helps construct prompts but does not retrieve data from external sources in real time.

A static knowledge base

Why wrong: A static knowledge base is not inherently integrated with real-time retrieval; it may be used but is not the essential component for RAG.

To detect outliers in the data

Why wrong: Outlier detection is not the purpose of normalization; it may even mask outliers.

To remove rows with missing values

Why wrong: Missing values are handled by the --drop-missing flag, not normalization.

To reduce the number of features from 30 to 2

Why wrong: Normalization does not reduce feature count; it only transforms values.

Create a custom scoring model specific to the business.

Why wrong: Standard model is used by default; custom is optional.

Deactivate any existing scoring rules.

Why wrong: Scoring rules can coexist with Einstein Opportunity Scoring.

Assign the Einstein Opportunity Scoring permission set to users.

Why wrong: This is needed for viewing scores, not for activation.

Ensuring the data set includes at least 100,000 records per prediction field.

Why wrong: While sample size matters, quality and relevance are more critical.

Creating a separate data warehouse to aggregate data from all objects.

Why wrong: Einstein works with Salesforce data directly; a data warehouse adds complexity.

Providing extensive user training on AI concepts.

Why wrong: User training helps adoption but not model accuracy.

Apply the strictest regulation globally.

Why wrong: While safe, it may over-restrict data use and not be efficient.

Store all data in a single data warehouse.

Why wrong: Centralizing storage does not address differing regulatory requirements.

Allow each region to manage its own data separately.

Why wrong: Decentralized management risks inconsistency and may complicate global AI initiatives.

Use a smaller file size limit.

Why wrong: While it might avoid the error, it doesn't solve the underlying memory problem and could lead to data loss.

Increase the number of worker threads.

Why wrong: More threads increase memory contention and may worsen the issue.

Switch to XML format.

Why wrong: XML is more verbose and would increase memory usage, not decrease.

Apply regularization.

Why wrong: Regularization helps with overfitting but is not specific to pair correlation.

Include both to capture more information.

Why wrong: High correlation can lead to unstable coefficient estimates.

Increase the sample size.

Why wrong: More data does not fix multicollinearity.

The dataset has an imbalanced class distribution

Why wrong: Imbalance typically causes poor recall on minority class, not a huge train/validation gap.

The dataset contains many missing values

Why wrong: Missing values often reduce overall performance, not specifically a gap.

The model was trained on stale data from a different season

Why wrong: Stale data would affect both sets similarly if validation is also outdated.