Professional Machine Learning Engineer Free Sample Questions

Predicting inventory from region, location, historical demand and seasonal popularity is a supervised forecasting problem over time-ordered data. Recurrent neural networks (for example, LSTMs) model temporal dependencies such as seasonality and trends, and the model can be retrained daily on the newest inventory data. Reinforcement learning learns a policy for taking actions from reward signals rather than predicting a labeled quantity, classification predicts discrete classes rather than a demand quantity, and CNNs are designed for spatial (image-like) data.

Low loss on the training data with worse performance on the validation data means that the model is overfitting. L2 regularization and dropout are the standard ways to reduce overfitting, and their best values are found by experiment. So run a hyperparameter tuning job on Agent Platform (formerly Vertex AI) that searches over the L2 regularization and dropout parameters. Choosing one fixed dropout or L2 value and reducing the learning rate is guesswork, and the learning rate isn't a regularization setting. Tuning only the learning rate and doubling the number of neurons adds model capacity, which makes overfitting worse.

Ticket priority and resolution time are organization-specific predictions that require custom models trained on your own ticket history, so they are served from Gemini Enterprise Agent Platform (formerly Vertex AI) endpoints. Sentiment analysis of general-language text with no domain-specific jargon is exactly what the pre-trained Cloud Natural Language API provides, so no custom sentiment model needs to be trained or maintained. Vision models do not apply to text tickets.

Fraud is the rare (less common) class. For AutoML tabular classification, the AUC PR objective "maximize the area under the precision-recall curve" optimizes results for inferences for the less common class, balancing precision (few false positives) and recall (catching fraud). AUC ROC is the binary default and weighs both classes, log loss optimizes probability calibration, and precision at a fixed recall of 0.50 would miss half of the fraud.

Dataflow can process more than a million calls per day at scale in the originating region and can call Sensitive Data Protection (DLP) to de-identify PII before storage. BigQuery provides regional datasets and GoogleSQL, an ANSI-compliant SQL dialect, for the third-party visualization tool. Pub/Sub is messaging rather than processing, Datastore has no ANSI SQL interface, and Cloud SQL and Cloud Run functions are not designed for analytics at this scale.

A Cloud Build trigger connected to your Git repository (for example, GitHub) starts a build only when new code is pushed, so retraining runs automatically (no manual intervention), only when the architecture actually changes (minimal compute cost), and the code is version-controlled in Git. Submitting jobs manually with gcloud requires intervention, Cloud Storage is not version control, and a daily Composer sensor adds cost and polling overhead.

In an Agent Platform Workbench instance the BigQuery client library is preinstalled, and the %%bigquery magic runs the SQL query, converts the results to a pandas DataFrame and can save it to a variable in one step (using the BigQuery Storage API for fast downloads). Exporting to CSV via Drive, local download or Cloud Storage adds unnecessary manual steps.

Google Cloud Pipeline Components provide a prebuilt BigqueryQueryJobOp that launches a BigQuery query job, waits for it to finish and exposes the result as a pipeline artifact for downstream steps. Reusing a prebuilt component is the easiest option; writing your own script or custom component adds code to build and maintain, and running the query manually in the console is not part of the automated pipeline.

Temperature readings are a time series. Readings an hour apart are highly correlated, so a random split puts near-copies of the test-period data into the training data. This data leakage made the test accuracy (97%) look much better than the real accuracy in production (66%). Google's guidance is to use a chronological split for time-series tasks. It also says to compute transformation statistics (for normalization or bucketizing) from the training split only and apply them to the validation and test splits. Applying transformations before splitting is itself a source of leakage. Adding more test data, or normalizing the training and test data as separate steps, doesn't remove the leakage that the random split causes.

A ResourceExhaustedError (OOM) means the tensors for one training step do not fit in GPU memory. Activation memory scales with the batch size, so reducing the batch size is the standard fix that keeps the model and input resolution unchanged. Changing the optimizer or learning rate does not materially change memory use, and reducing the image shape would change the model input and can hurt accuracy.