CompTIA DataX DY0-001 (V1) Practice Question
A payment-processing platform is evaluating a gradient-boosted decision-tree (GBDT) fraud-detection model against the company's long-standing rule-based filter. During a 30-day A/B test the following aggregated results were collected:
Metric Rule-based | GBDT
--------------------------------------------------------------
Precision 0.52 | 0.71
Recall 0.78 | 0.80
F1 score 0.62 | 0.75
Average inference latency (ms) 22 | 48
False positives per million transactions 9600 | 4800
False negatives per million transactions 2600 | 2400
Infrastructure cost per 1 M inferences (USD) 25 | 60
The service-level agreement (SLA) requires latency to stay below 75 ms and fewer than 6000 false positives per million transactions. Monthly volume is 40 million transactions. Finance estimates that each false positive costs USD 2.50 in manual-review labor, while each false negative leads to an average chargeback loss of USD 15.
Which statement most strongly justifies recommending the GBDT model over the conventional rule-based process?
Because confidence intervals for precision and recall were not reported, the results cannot justify replacing the well-understood rule-based system.
The GBDT satisfies all SLA limits and, after accounting for error-related costs, is projected to save roughly USD 600 000 per month even after its extra USD 1 400 infrastructure bill.
Although the GBDT improves F1, its inference latency more than doubles, so user-experience risk outweighs any potential savings.
The infrastructure cost of the GBDT increases by 140 %, making it economically infeasible despite its modest gain in recall.