Your company manages a video sharing website where users can watch and upload videos. You need to create an ML model to predict which newly uploaded videos will be the most popular so that those videos can be prioritized on your company's website. Which result should you use to determine whether the model is successful?
Answer(s): C
Option C is correct because it aligns evaluation with business objective: predicting the proportion of videos that become popular as measured by watch time within 30 days directly reflects user engagement and long-term value, which is a typical metric for popularity prediction in video platforms.A) Incorrect — using uploader likes is a proxy signal and circular bias; it does not measure actual viewer engagement or video performance after upload.B) Incorrect — 97.5% of clickbait videos by number of clicks focuses on short-term clicks, not sustained watch time, and can incentivize low-quality content.D) Incorrect — a Pearson correlation of 0 indicates no linear relationship, not a meaningful measure of predictive success.
You are working on a Neural Network-based project. The dataset provided to you has columns with different ranges. While preparing the data for model training, you discover that gradient optimization is having difficulty moving weights to a good solution. What should you do?
Answer(s): B
Option B is correct because normalization (representation transformation) scales features to a similar range, which stabilizes and accelerates gradient-based optimization in neural networks, helping weights converge more effectively.A) Incorrect — feature construction may help if features are weak, but it does not directly address gradient optimization stability or scaling issues.C) Incorrect — removing features with missing values addresses data quality, not gradient optimization or feature scale; imputation or robust modeling would be more relevant.D) Incorrect — partitioning to alter train/test sizes does not fix feature scaling or optimization dynamics and can harm evaluation validity.
Your data science team needs to rapidly experiment with various features, model architectures, and hyperparameters. They need to track the accuracy metrics for various experiments and use an API to query the metrics over time. What should they use to track and report their experiments while minimizing manual effort?
Answer(s): A
Option A is correct because Kubeflow Pipelines provides a robust framework to orchestrate ML experiments and exposes an API to access run-level metrics and lineage, enabling rapid experimentation with minimal manual effort.B is incorrect because while AI Platform Training can run experiments, exporting metrics to BigQuery and querying via the BigQuery API adds unnecessary complexity for time-series experiment metrics tracking and does not leverage a purpose-built ML experiment-tracking workflow.C is incorrect because Cloud Monitoring is designed for infrastructure and service metrics, not experiment-level ML metrics, and using its API is not the optimal path for tracking model accuracy across experiments.D is incorrect because AI Platform Notebooks with Google Sheets introduces manual data collection and lacks scalable, programmatic experiment tracking and versioning that a dedicated ML pipeline provides.
You work for a bank and are building a random forest model for fraud detection. You have a dataset that includes transactions, of which 1% are identified as fraudulent. Which data transformation strategy would likely improve the performance of your classifier?
Option C is correct because oversampling the minority class (fraud) addresses severe class imbalance, improving the model’s ability to learn decision boundaries for fraud cases and typically boosting metrics like recall/precision in imbalanced fraud detection with tree-based models.A) Incorrect — Writing data in TFRecords is an I/O/performance optimization, not a predictive feature transformation that directly improves classifier performance.B) Incorrect — Z-normalization is not always beneficial for tree ensembles like random forests, which are tree-based and less sensitive to feature scaling.D) Incorrect — One-hot encoding can increase sparsity and dimensionality without addressing class imbalance, potentially harming model efficiency and performance.
https://towardsdatascience.com/how-to-build-a-machine-learning-model-to-identify-credit-card- fraud-in-5-stepsa-hands-on-modeling-5140b3bd19f1
You are using transfer learning to train an image classifier based on a pre-trained EfficientNet model. Your training dataset has 20,000 images. You plan to retrain the model once per day. You need to minimize the cost of infrastructure. What platform components and configuration environment should you use?
Answer(s): D
Option D is correct because AI Platform Training with a custom scale tier allows on-demand, scalable training jobs (4 V100 GPUs) using Cloud Storage for data and artifacts, minimizing idle infrastructure and cost while supporting daily retraining.A) Incorrect — 4 V100 GPUs with local storage incurs higher cost and doesn’t leverage scalable, managed training; local storage isn’t optimal for distributed training or long-term cost efficiency.B) Incorrect — 4 V100 GPUs with Cloud Storage still uses a fixed VM setup; no managed scaling or orchestration, increasing management overhead and cost for daily retraining.C) Incorrect — GKE with an NFS server adds complexity and persistent storage management; not optimized for cost-efficient, single-job daily retraining.
While conducting an exploratory analysis of a dataset, you discover that categorical feature A has substantial predictive power, but it is sometimes missing. What should you do?
Option D is correct because treating missingness as a signal by adding an indicator for missing values and keeping A's original category leverages the informative nature of missingness and preserves information for models that can handle categorical features. A) Incorrect — arbitrarily dropping based on a threshold ignores the potential predictive value of A and the missingness signal. B) Incorrect — imputing with mode can bias category distributions and ignore missingness as information. C) Incorrect — imputing from the single most correlated feature risks leakage and ignores the explicit missingness pattern.
You work for a large retailer and have been asked to segment your customers by their purchasing habits. The purchase history of all customers has been uploaded to BigQuery. You suspect that there may be several distinct customer segments, however you are unsure of how many, and you don't yet understand the commonalities in their behavior. You want to find the most efficient solution. What should you do?
Option A is correct because k-means clustering with BigQuery ML can perform unsupervised segmentation directly on the existing BigQuery data, and BigQuery ML can automatically determine a reasonable number of clusters (e.g., via elbow method or in-model initialization), making it efficient for discovering customer segments without labeled data.B is incorrect because Dataprep is for data cleaning/prep, not for discovering latent segments or clustering similarities across columns; it doesn’t perform unsupervised segmentation.C is incorrect because Data Labeling Service and AutoML Tables are for supervised learning with labeled data; without labeled segments, this isn’t appropriate for discovering inherent purchase-based segments.D is incorrect because labeling to a fixed list from Marketing introduces predefined categories; it does not discover natural segments and adds labeling overhead.
You recently designed and built a custom neural network that uses critical dependencies specific to your organization's framework. You need to train the model using a managed training service on Google Cloud. However, the ML framework and related dependencies are not supported by AI Platform Training. Also, both your model and your data are too large to fit in memory on a single machine. Your ML framework of choice uses the scheduler, workers, and servers distribution structure. What should you do?
Option C is correct because AI Platform Training supports custom containers for distributed training, enabling large-scale training with the framework’s scheduler, workers, and servers architecture when native support is lacking. The scenario requires distributed training across multiple machines, which is accommodated by custom containers on AI Platform Training. A) is incorrect because built-in models don’t address the need for your custom, large-scale, framework-based training. B) insufficient because single-container jobs don’t cover distributed training across multiple workers/servers. D) unnecessary since you can use distributed custom containers without changing to a different framework.
Share your comments for Google PROFESSIONAL MACHINE LEARNING ENGINEER exam with other users:
good ................
Keeping this site free takes real effort. We constantly battle automated scraping and unauthorized content copying. A quick account helps us protect the community and keep the site free.
To continue studying for your PROFESSIONAL MACHINE LEARNING ENGINEER, please sign in or create a free account.