A confusion matrix (rows actual 0/1, columns predicted 0/1) is [[11, 1], [2, 6]]. What is the recall?
Q2
For the "station will be empty" model, what does precision mean operationally?
Q3
Overall recall is 0.75, but recall is 1.00 for central stations and 0.50 for peripheral ones. What does this show?
Q4
What does an ROC AUC of 0.72 mean?
Q5
Why are precision–recall curves often preferred over ROC curves when positives are rare (e.g., 1 % of stations empty)?
Q6
Lowering a classifier's threshold usually:
Q7
Who should choose the decision threshold of the empty-station model, and with which visualization?
Q8
A reliability diagram shows that stations predicted at "80 % chance of becoming empty" actually become empty only 50 % of the time. The model is:
Q9
What does permutation importance measure?
Q10
A SHAP summary shows "station id" as the most important feature of the demand model. What is the right interpretation?
Q11
What does a partial dependence plot of predicted trips vs. temperature show?
Q12
What is data drift?
Q13
A feature's population stability index is 0.13. By the common rule of thumb, this indicates:
Q14
A model trained on winter data is used in summer; temperature PSI is 5.9. What should the monitoring dashboard lead to?
Q15
Request latencies have a mean of 142 ms and a 95th percentile of 371 ms. Which statement is correct?
Q16
What is the difference between tumbling and sliding windows in stream processing?
Q17
An alert fires when p95 latency > 200 ms and clears when it drops below 200 ms. With p95 hovering around 200 ms, what happens, and what is the fix?
Q18
A service-level objective is "99.9 % of requests under 200 ms per month". What is the error budget?
Q19
Which pairing is typical of an open-source observability stack?
Q20
In an edge–cloud system, a phone's request goes to an edge server, which offloads heavy work to the cloud. Why should the latency dashboard decompose end-to-end latency?
Q21
Training and validation loss curves are plotted per epoch. Training loss keeps falling while validation loss starts rising after epoch 12. What does this suggest?
Q22
A model has 98 % accuracy at predicting empty stations, but only 2 % of station-hours are actually empty. Why is accuracy misleading here?
Q23
Which chart directly shows whether a model's predicted probabilities can be taken at face value?
Q24
Why do dashboards for live services show p95/p99 latency rather than only the mean?
Q25
An error-budget burn-rate alert fires when:
Q26
Reference data has values 1–8; current data is the same values shifted by +1 (2–9). Using 4 quantile bins of the reference, the PSI is about 0.14. How should the dashboard label this?
Q27
For a fairness review of the empty-station model, which visualization is most useful first?
Q28Short answer
Using the scores and labels: positives scored 0.95, 0.85, 0.70, 0.45, 0.15; negatives scored 0.80, 0.55, 0.30, 0.20, 0.05. (a) Compute the ROC AUC by counting pairs. (b) Give recall and precision at thresholds 0.70 and 0.45. (c) Which threshold would you recommend if missing an empty station is three times as costly as an unnecessary van trip?
Q29Short answer
Design the monitoring dashboard for CityPulse's deployed demand-prediction model. Name the panels, the statistic and comparison in each, the alert rules (with persistence and hysteresis), and who uses which panel.
Q30Short answer
For an edge–cloud video-analytics system (cameras → edge servers → cloud), propose a dashboard that helps engineers decide when to offload more work to the cloud. Specify the metrics per tier, the idioms, and how you would show the trade-off between latency, edge load and cloud cost.