rohitg00--ai-engineering-from-scratch
wehub-resource-sync
68 行
3.4 KiB
JSON
68 行
3.4 KiB
JSON
[
|
|
{
|
|
"id": "anomaly-pre-1",
|
|
"stage": "pre",
|
|
"question": "Why is anomaly detection typically framed as an unsupervised problem rather than classification?",
|
|
"options": [
|
|
"Anomaly detection does not require any data",
|
|
"Labeled anomalies are extremely rare, and novel anomaly types differ from previously seen ones",
|
|
"Supervised classification is always less accurate",
|
|
"Anomaly detection only works on time series data"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Anomalies are rare (often <0.1% of data), so there are too few labeled examples to train a classifier. Also, future anomalies may be of types never seen before. Modeling 'normal' and flagging deviations handles both problems."
|
|
},
|
|
{
|
|
"id": "anomaly-pre-2",
|
|
"stage": "pre",
|
|
"question": "A temperature of 90F is normal in summer but anomalous in winter. What type of anomaly is this?",
|
|
"options": [
|
|
"Point anomaly",
|
|
"Contextual anomaly",
|
|
"Collective anomaly",
|
|
"Statistical anomaly"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "A contextual anomaly is a value that is unusual given its context (time, location). The value itself might be normal in a different context. Same data point, different interpretation based on surrounding conditions."
|
|
},
|
|
{
|
|
"id": "anomaly-post-1",
|
|
"stage": "post",
|
|
"question": "The Z-score method flags points more than 3 standard deviations from the mean. When does this approach fail?",
|
|
"options": [
|
|
"When the data is perfectly normally distributed",
|
|
"When the data is multimodal, skewed, or when outliers in the training data inflate the mean and std",
|
|
"When there are exactly 3 anomalies in the dataset",
|
|
"When the features are standardized"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Z-score assumes a single Gaussian distribution. It fails on multimodal data (multiple clusters), skewed distributions, and when outliers in training data shift the mean/std, making real anomalies harder to detect."
|
|
},
|
|
{
|
|
"id": "anomaly-post-2",
|
|
"stage": "post",
|
|
"question": "How does Isolation Forest detect anomalies differently from distance-based methods?",
|
|
"options": [
|
|
"It uses neural networks instead of trees",
|
|
"It isolates points using random splits; anomalies require fewer splits to isolate because they are few and different",
|
|
"It computes distances to every other point in the dataset",
|
|
"It only works on text data"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Isolation Forest randomly partitions data with tree splits. Anomalies, being rare and different, end up isolated (in their own leaf) with fewer splits. Short average path length = more anomalous."
|
|
},
|
|
{
|
|
"id": "anomaly-post-3",
|
|
"stage": "post",
|
|
"question": "You build both an unsupervised anomaly detector and a supervised fraud classifier. When should you prefer the unsupervised approach?",
|
|
"options": [
|
|
"Always -- unsupervised is always better for anomaly detection",
|
|
"When you need to detect novel fraud patterns that differ from historical labeled examples",
|
|
"When you have millions of labeled fraud examples",
|
|
"When you only care about precision, not recall"
|
|
],
|
|
"correct": 1,
|
|
"explanation": "Supervised classifiers only catch fraud types present in training data. Unsupervised detectors flag any deviation from normal, catching novel fraud schemes. The tradeoff is higher false positive rate."
|
|
}
|
|
]
|