Fundamentals of AI and ML
Free AWS Certified AI Practitioner (AIF-C01) practice — 6 questions on Fundamentals of AI and ML, with explanations. No sign-up. Full 12-question mixed test →
A healthcare startup is building a system to predict patient readmission risk. They have 500,000 patient records with 45 features each, but only 3% of patients were readmitted. Initial model testing shows 97% accuracy, but the business team reports the model fails to identify most at-risk patients. What is the PRIMARY issue and appropriate solution?
With only 3% positive cases, a model that always predicts 'no readmission' achieves 97% accuracy while being completely useless. Class imbalance is the core issue. The business complaint ('fails to identify at-risk patients') confirms low recall on the minority class. Solutions include resampling techniques (SMOTE), adjusting class weights in the loss function, or optimizing for F1-score/AUC-ROC instead of accuracy.
An e-commerce company wants to reduce the dimensionality of their product catalog dataset containing 200 features (price, ratings, category embeddings, user interaction counts) down to 20 features for clustering analysis. They need to interpret which original features most influence each reduced dimension for business insights. Which technique should they use?
PCA provides feature loadings (coefficients) for each principal component, showing exactly how much each original feature contributes to each new dimension. This interpretability is critical for the business requirement to understand 'which original features most influence each reduced dimension.' PCA components are linear combinations, making them mathematically traceable back to original features.
A financial institution is developing a loan approval model. Regulatory requirements mandate that the institution must be able to explain each individual loan decision to customers and auditors, including which specific factors led to approval or denial. The dataset has 50,000 applications with 80 features including credit score, income, debt ratios, and employment history. Which modeling approach BEST satisfies this requirement?
Logistic regression provides inherent interpretability through coefficients that directly show how each feature influences the log-odds of approval. For regulatory compliance, you can state: 'A 1-unit increase in credit score increases approval probability by X%.' The model is deterministic, auditable, and the explanations are part of the model itself (not post-hoc approximations). This level of transparency is critical for regulated industries.
A data scientist is training a CNN for medical image classification. During training, they observe: Training loss: 0.15, Validation loss: 0.14, Test accuracy: 94%, Validation accuracy: 93%. However, when deployed to a hospital using images from a different scanner model, accuracy drops to 78%. What is the MOST likely root cause?
The key diagnostic clue is that training/validation performance are nearly identical (no overfitting), and test set performance is good (93-94%), but deployment performance dramatically drops to 78%. This indicates the test set was drawn from the same distribution as training data, but real-world deployment data (different scanner) has a different distribution. This is classic distribution/domain shift - the model learned features specific to one scanner type that don't transfer to another.
A recommendation system uses collaborative filtering to suggest products. The user-item interaction matrix is 1,000,000 users × 500,000 items, but only 0.01% of entries are non-zero (extremely sparse). Matrix factorization via SVD is computationally prohibitive. Which approach is MOST appropriate for this scale and sparsity?
ALS with mini-batch gradient descent is specifically designed for sparse matrix factorization at scale. It only processes observed (non-zero) entries, making it computationally feasible. ALS alternates between fixing user factors and solving for item factors (and vice versa), which is more efficient than full SVD. This is the standard approach for large-scale recommender systems (used by Spotify, Netflix).
A time-series forecasting model predicts hourly energy consumption. The data scientist splits data as: records 1-10,000 for training, records 10,001-12,000 for validation, records 12,001-14,000 for testing. After deployment, forecast accuracy is significantly worse than test metrics. What is the MOST critical flaw in the evaluation approach?
Time-series data often exhibits non-stationarity - patterns change over time due to seasonality shifts, trend changes, or external factors. The test set (records 12,001-14,000) may have been from a specific historical period that doesn't represent current deployment conditions. The key clue is 'after deployment' performance drops, suggesting the most recent data (post-14,000) has different characteristics than the historical test set. Proper evaluation should use the most recent data as the test set.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$54.99 $17.99 with code FREETEST33 — valid through August 22.
Get my $17.99 deal →