[
  {
    "id": "knn-pre-1",
    "stage": "pre",
    "question": "KNN is called a 'lazy learner.' What does that mean?",
    "options": [
      "It converges slowly during training",
      "It uses a simplified version of the loss function",
      "It does no computation at training time and computes everything at prediction time",
      "It only works on small datasets"
    ],
    "correct": 2,
    "explanation": "Lazy learning means KNN stores the training data and does no work during 'training.' All computation (distance calculations, voting) happens when a prediction is requested."
  },
  {
    "id": "knn-pre-2",
    "stage": "pre",
    "question": "Why is feature scaling critical for KNN?",
    "options": [
      "KNN cannot handle negative numbers without scaling",
      "KNN uses gradient descent which requires normalized inputs",
      "Feature scaling reduces the number of neighbors needed",
      "Distance calculations are dominated by features with larger ranges, making scaling necessary for fair comparisons"
    ],
    "correct": 3,
    "explanation": "KNN relies on distances. A feature ranging 0-1000 will dominate a feature ranging 0-1 in distance calculations. Scaling puts all features on comparable ranges."
  },
  {
    "id": "knn-post-1",
    "stage": "post",
    "question": "In 100 dimensions with uniform random points, what happens to the ratio of max distance to min distance?",
    "options": [
      "It increases dramatically, making neighbors more distinct",
      "It approaches 1, making all points nearly equidistant from each other",
      "It stays the same as in 2 dimensions",
      "It becomes negative due to numerical overflow"
    ],
    "correct": 1,
    "explanation": "This is the curse of dimensionality. In high dimensions, distances converge: max_dist / min_dist approaches 1. When all points are equidistant, 'nearest' becomes meaningless."
  },
  {
    "id": "knn-post-2",
    "stage": "post",
    "question": "Which distance metric is most appropriate for comparing text documents represented as TF-IDF vectors?",
    "options": [
      "Chebyshev distance",
      "Cosine distance",
      "L2 (Euclidean) distance",
      "L1 (Manhattan) distance"
    ],
    "correct": 1,
    "explanation": "Cosine distance measures the angle between vectors, ignoring magnitude. For text, document length (magnitude) is noise -- direction captures meaning. Cosine consistently outperforms L1/L2 for text."
  },
  {
    "id": "knn-post-3",
    "stage": "post",
    "question": "What happens to the KNN decision boundary as K increases from 1 to N (the full dataset size)?",
    "options": [
      "The boundary becomes more complex and detailed",
      "The boundary stays the same regardless of K",
      "The boundary becomes circular",
      "The boundary smooths out, eventually predicting the majority class for every point"
    ],
    "correct": 3,
    "explanation": "K=1 produces jagged boundaries that follow every point (overfitting). As K increases, boundaries smooth out. K=N means every query considers all points, always predicting the majority class (maximum bias)."
  }
]