Real-Time Fraud Detection with AutoML Tabular and Managed Endpoints
Short answer
D Train an Agent Platform AutoML model on the BigQuery tabular transaction data and deploy it to an Agent Platform endpoint for managed, low-latency real-time fraud scoring.
Question
Your team is proficient in data science but wants to avoid the complexity of managing and optimizing a real-time model serving infrastructure. What should you do?
Key Takeaways
It tests distinguishing managed online prediction from batch or forecasting approaches when hard per-transaction latency and scale are required.
This PMLE scenario tests real-time fraud detection on BigQuery tabular data under a 200 ms latency target and high daily volume. The correct design is Agent Platform AutoML tabular training deployed to a managed endpoint (D), avoiding self-managed serving infrastructure.
Common mistake: The most plausible wrong answer is B: using BigQuery ML `ML.PREDICT` in a per-minute batch job, which looks simple and BigQuery-native but cannot return a prediction for each transaction within 200 ms.
Expert Analysis
Real-time is the gating qualifier
The question asks for a per-transaction approve or deny decision in under 200 ms. That rules out any design whose smallest unit is a scheduled batch: option B uses BigQuery ML ML.PREDICT every minute to score recent transactions after the fact, so it cannot meet the stated per-request latency even if aggregate daily throughput is fine.
Managed serving removes the stated burden The team explicitly wants to avoid managing and optimizing real-time serving infrastructure. Option D uses Agent Platform AutoML on the BigQuery tabular dataset and deploys to an Agent Platform endpoint, which supplies a managed API and autoscaling. This directly satisfies both the low-latency online prediction and millions-per-day scale requirements.
Treat the task as classification, not forecasting Option C uses Forecasting with AutoML, a time-series method. Fraud detection here is transaction-level binary classification, not predicting which future time windows contain fraud. Forecasting would not produce the immediate approve or deny response the application needs, so C is misaligned regardless of latency.
Reject the self-managed GPU path Option A can train a DNN in Workbench and serve it, but it requires a GPU-enabled Compute Engine cluster plus a custom API. That is precisely the serving infrastructure management the team wants to avoid, and the added operational complexity and cost are unnecessary for structured tabular data where AutoML is appropriate.
Why D wins D is the only option that combines BigQuery-native tabular training with a managed online endpoint. It preserves real-time inference, scales automatically, and minimizes serving operations; the correct answer is D.
Official Reference
- https://cloud.google.com/vertex-ai/docs/tabular-data/classification-regression/overview
- https://cloud.google.com/vertex-ai/docs/predictions/overview
- https://cloud.google.com/bigquery-ml/docs/reference/standard-sql/bigqueryml-syntax-predict
These links open the official vendor documentation in a new tab.
Exam Strategy
When a stem specifies a hard per-request latency and high throughput, first mark every scheduled-batch design as disqualified; batch cadence is not online serving. Then scan for the phrase about avoiding infrastructure management: managed endpoints or serverless prediction services are strong signals, while custom GPU clusters and self-built APIs lose. Finally, verify the ML task matches the label—per-transaction fraud is classification, not time-series forecasting.
Frequently Asked Questions
Why can't BigQuery ML batch prediction meet the 200 ms requirement?
ML.PREDICT job scores transactions in batches after they arrive; it does not answer each authorization request synchronously. The stated under-200 ms decision must come from an online endpoint.What makes Agent Platform AutoML suitable for the BigQuery tabular fraud dataset?
Why is Forecasting with AutoML wrong for this fraud use case?
Community Votes
100% of anonymous learners picked answer D. Votes are pick records left by other test-takers — they are not the verified answer.
Discussion 2 contributions
Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.
From public community threads 2
-
taylz876 👍 1 Answered DSimilar to Q240 and Q272 Why D is right • Real-time with low latency: a deployed endpoint answers each transaction immediately, within the 200 ms limit. • Scales without managing infrastructure: the endpoint is managed and autoscales to millions of transactions a day. • Tabular data in BigQuery, a team that wants to avoid complexity: AutoML trains on the tabular data with little code and no serving infrastructure to build. Why the others are wrong • A: A DNN on GPU-enabled Compute Engine instances with a custom API is the infrastructure complexity the team wants to avoid. • B: Scoring every minute with ML.PREDICT is batch, not a prediction for each transaction as it happens. • C: A forecasting model on sequences predicts future time windows, not whether one transaction is fraud right now.
-
n2183712847 👍 1 Answered DCorrect answer: D. Agent Platform AutoML can train directly on tabular data from BigQuery and deploy the model to a managed endpoint that automatically handles scalable real-time serving with low latency; A requires substantial infrastructure management, B uses batch prediction and cannot reliably meet a sub-200 ms per-transaction requirement, and C is inappropriate because fraud detection is a classification problem rather than time-series forecasting.
Reader comments & corrections
No reader corrections yet — spotted an error or have a note? Share it below.
Related Analysis
Practice All PMLE Questions
Access 65 questions with complete answers and detailed explanations.
View Full PMLE Practice Test →