Testing for Algorithmic, Sample and Inappropriate Bias CT-AI Practice Questions (Page 4)

Questions 16–20

1. 16expert · hard · select all that apply

   A team is testing a speech recognition system for a virtual assistant. The training data was collected from users who opted into a beta program, which is known to be skewed toward younger, tech-savvy users. The team suspects sample bias. Select all that apply: Which testing techniques would help confirm and assess the sample bias?

   • AEvaluate the model's accuracy on a test set that includes a diverse range of ages and accents.
   • BCollect additional training data from underrepresented age groups and retrain the model.
   • CCompare the age distribution in the training data with the general population.
   • DRun a fairness audit comparing error rates across age groups.
   • EReview the model's architecture to ensure it uses appropriate acoustic models.
   Check answers Discussion Select an answer first
2. 17foundation · easy

   What is the primary cause of sample bias in a machine learning model?

   • AThe training data does not accurately represent the population or scenarios the model will encounter in production.
   • BThe model's algorithm is too complex and overfits the training data.
   • CThe test data is labeled incorrectly by human annotators.
   • DThe model's objective function prioritizes accuracy over fairness.
   Reveal Solution Discussion Select an answer first
4. 18expert · hard

   A team is testing a predictive policing model. The training data includes historical crime reports, which are known to reflect over-policing in certain neighborhoods. The model is being evaluated for bias. The team is considering several testing techniques. Which combination of techniques would be most effective in identifying both sample bias and algorithmic bias?

   • AReview the model's code for hardcoded rules and test with a synthetic dataset.
   • BAnalyze the training data for representativeness and run a fairness audit on the model's predictions across neighborhoods.
   • CCompare the model's accuracy across neighborhoods and retrain with more data from over-policed areas.
   • DUse a single fairness metric, such as demographic parity, to assess the model.
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   Which testing approach is best for assessing inappropriate bias in an AI system's outcomes?

   • AChecking the model's training data for missing values.
   • BMeasuring the model's inference speed on different hardware.
   • CEvaluating the model's predictions across different demographic groups to identify systematic disparities.
   • DReviewing the model's documentation for compliance with GDPR.
   Reveal Solution Discussion Select an answer first
6. 20application · medium

   A retail company is building a recommendation system. The data science team collects data from users who have opted into the loyalty program. They notice that the model's recommendations are less relevant for users who do not shop frequently. The team suspects that the training data over-represents frequent shoppers. Which source of bias is most likely, and what testing approach would help confirm it?

   • AAlgorithmic bias; test the model with a synthetic dataset that has equal numbers of frequent and infrequent shoppers.
   • BSample bias; compare the distribution of shopping frequency in the training data with the overall customer population.
   • CInappropriate bias; run a fairness audit on recommendation relevance across different age groups.
   • DSample bias; increase the model's complexity to better fit the infrequent shoppers.
   Reveal Solution Discussion Select an answer first