Google Professional Machine Learning Engineer Practice Test - Questions Answers, Page 7

Data leakage is a problem where information from outside the training dataset is used to create the model, resulting in an overly optimistic or invalid estimate of the model performance. Data leakage can occur in time series data when the temporal order of the data is not preserved during data preparation or model evaluation. For example, if the data is shuffled before splitting into train and test sets, or if future data is used to impute missing values in past data, then data leakage can occur.

One way to address data leakage in time series data is to apply nested cross-validation during model training. Nested cross-validation is a technique that allows you to perform both model selection and model evaluation in a robust way, while preserving the temporal order of the data. Nested cross-validation involves two levels of cross-validation: an inner loop for model selection and an outer loop for model evaluation. The inner loop splits the training data into k folds, trains and tunes the model on k-1 folds, and validates the model on the remaining fold. The inner loop repeats this process for each fold and selects the best model based on the validation performance. The outer loop splits the data into n folds, trains the best model from the inner loop on n-1 folds, and tests the model on the remaining fold. The outer loop repeats this process for each fold and evaluates the model performance based on the test results.

Nested cross-validation can help to avoid data leakage in time series data by ensuring that the model is trained and tested on non-overlapping data, and that the data used for validation is never seen by the model during training. Nested cross-validation can also provide a more reliable estimate of the model performance than a single train-test split or a simple cross-validation, as it reduces the variance and bias of the estimate.

Data Leakage in Machine Learning

How to Avoid Data Leakage When Performing Data Preparation

Classification on a single time series - prevent leakage between train and test

In this scenario, the goal is to predict the most relevant web banner that a user should see next on an online travel agency's website. The model needs to have low latency requirements of 300ms@p99, and there are thousands of web banners to choose from. The exploratory analysis has shown that the navigation context is a good predictor. Security is also important to the company. Given these requirements, the best configuration for the prediction pipeline would be to embed the client on the website and deploy the model on AI Platform Prediction. Option A is the correct answer.

Option A: Embed the client on the website, and then deploy the model on AI Platform Prediction. This option is the simplest solution that meets the requirements. The client can collect the user's navigation context and send it to the model deployed on AI Platform Prediction for prediction. AI Platform Prediction can handle large-scale prediction requests and has low latency requirements. This option does not require any additional infrastructure or services, making it the simplest solution.

Option B: Embed the client on the website, deploy the gateway on App Engine, and then deploy the model on AI Platform Prediction. This option adds an additional layer of infrastructure by deploying the gateway on App Engine. While App Engine can handle large-scale requests, it adds complexity to the pipeline and may not be necessary for this use case.

Option C: Embed the client on the website, deploy the gateway on App Engine, deploy the database on Cloud Bigtable for writing and for reading the user's navigation context, and then deploy the model on AI Platform Prediction. This option adds even more complexity to the pipeline by deploying the database on Cloud Bigtable. While Cloud Bigtable can provide fast and scalable access to the user's navigation context, it may not be needed for this use case. Moreover, Cloud Bigtable may introduce additional latency and cost to the pipeline.

Option D: Embed the client on the website, deploy the gateway on App Engine, deploy the database on Memorystore for writing and for reading the user's navigation context, and then deploy the model on Google Kubernetes Engine. This option is the most complex and costly solution that does not meet the requirements. Deploying the model on Google Kubernetes Engine requires more management and configuration than AI Platform Prediction. Moreover, Google Kubernetes Engine may not be able to meet the low latency requirements of 300ms@p99. Deploying the database on Memorystore also adds unnecessary overhead and cost to the pipeline.

AI Platform Prediction documentation

App Engine documentation

Cloud Bigtable documentation

[Memorystore documentation]

[Google Kubernetes Engine documentation]

In this scenario, the goal is to speed up model training for a CNN-based architecture on Google Cloud. The code does not include any manual device placement and has not been wrapped in Estimator model-level abstraction. Given these constraints, the best environment to train the model on would be a Deep Learning VM with an n1-standard-2 machine and 1 GPU with all libraries pre-installed. Option C is the correct answer.

Option C: A Deep Learning VM with an n1-standard-2 machine and 1 GPU with all libraries pre-installed. This option is the most suitable for the scenario because it provides a ready-to-use environment for deep learning on Google Cloud. A Deep Learning VM is a specialized VM image that is pre-installed with popular deep learning frameworks such as TensorFlow, PyTorch, Keras, and more. A Deep Learning VM also comes with NVIDIA GPU drivers and CUDA libraries that enable GPU acceleration for model training. A Deep Learning VM can be easily configured and launched from the Google Cloud Console or the Cloud SDK. An n1-standard-2 machine is a general-purpose machine type that provides 2 vCPUs and 7.5 GB of memory. This machine type can be sufficient for running a CNN-based architecture. A GPU is a specialized hardware accelerator that can speed up the computation of matrix operations and convolutions, which are common in CNN-based architectures. By using a Deep Learning VM with an n1-standard-2 machine and 1 GPU, the model training can be significantly faster than on an on-premises CPU-only infrastructure.

Option A: A VM on Compute Engine and 1 TPU with all dependencies installed manually. This option is not suitable for the scenario because it requires manual installation of dependencies and device placement. A TPU is a custom-designed ASIC that can provide high performance and efficiency for TensorFlow models. However, to use a TPU, the code needs to include manual device placement and be wrapped in Estimator model-level abstraction. Moreover, to use a TPU, the dependencies such as TensorFlow, Cloud TPU Client, and Cloud Storage need to be installed manually on the VM. This option can be complex and time-consuming to set up and may not be compatible with the existing code.

Option B: A VM on Compute Engine and 8 GPUs with all dependencies installed manually. This option is not suitable for the scenario because it requires manual installation of dependencies and may not be cost-effective. While using 8 GPUs can provide high parallelism and speed for model training, it also increases the cost and complexity of the environment. Moreover, to use GPUs, the dependencies such as NVIDIA GPU drivers, CUDA libraries, and deep learning frameworks need to be installed manually on the VM. This option can be tedious and error-prone to set up and may not be necessary for the scenario.

Option D: A Deep Learning VM with more powerful CPU e2-highcpu-16 machines with all libraries pre-installed. This option is not suitable for the scenario because it does not leverage GPU acceleration for model training. While using more powerful CPU machines can provide more compute resources and memory for model training, it may not be as fast and efficient as using GPU machines. CPU machines are not optimized for matrix operations and convolutions, which are common in CNN-based architectures. Moreover, using more powerful CPU machines can also increase the cost of the environment. This option can be suboptimal and wasteful for the scenario.

Deep Learning VM Image documentation

Compute Engine documentation

Cloud TPU documentation

Machine types documentation

GPUs on Compute Engine documentation

Labels are key-value pairs that you can attach to AI Platform resources such as jobs, models, and versions. Labels can help you organize your resources into descriptive categories that reflect your business needs. For example, you can use labels to indicate the owner, purpose, environment, or status of a resource. You can also use labels to filter the results when you list or monitor your resources on the Google Cloud Console or the Cloud SDK. Using labels can help you manage your resources in a clean and scalable way, without requiring separate projects or restrictive permissions.

Using labels to organize AI Platform resources

Creating and managing labels

In this scenario, the goal is to create a custom fraud detection model using AutoML Tables. Fraud detection is a type of binary classification problem, where the model needs to predict whether a transaction is fraudulent or not. The optimization objective is a metric that defines how the model is trained and evaluated. AutoML Tables allows you to choose from different optimization objectives for binary classification problems, such as Log loss, Precision at a Recall value, AUC PR, and AUC ROC.

To choose the best optimization objective for fraud detection, we need to consider the characteristics of the problem and the data. Fraud detection is a problem where the positive class (fraudulent transactions) is very rare compared to the negative class (legitimate transactions). This means that the data is highly imbalanced, and the model needs to be sensitive to the minority class. Moreover, fraud detection is a problem where the cost of false negatives (missing a fraudulent transaction) is much higher than the cost of false positives (flagging a legitimate transaction as fraudulent). This means that the model needs to have high recall (the ability to detect all fraudulent transactions) while maintaining high precision (the ability to avoid false alarms).

Given these considerations, the best optimization objective for fraud detection is the one that maximizes the area under the precision-recall curve (AUC PR) value. The AUC PR value is a metric that measures the trade-off between precision and recall for different probability thresholds. A higher AUC PR value means that the model can achieve high precision and high recall at the same time. The AUC PR value is also more suitable for imbalanced data than the AUC ROC value, which measures the trade-off between the true positive rate and the false positive rate. The AUC ROC value can be misleading for imbalanced data, as it can give a high score even if the model has low recall or low precision.

Therefore, option C is the correct answer. Option A is not suitable, as Log loss is a metric that measures the difference between the predicted probabilities and the actual labels, and does not account for the trade-off between precision and recall. Option B is not suitable, as Precision at a Recall value is a metric that measures the precision at a fixed recall level, and does not account for the trade-off between precision and recall at different thresholds. Option D is not suitable, as AUC ROC is a metric that can be misleading for imbalanced data, as explained above.

AutoML Tables documentation

Optimization objectives for binary classification

Precision-Recall Curves: How to Easily Evaluate Machine Learning Models in No Time

ROC Curves and Area Under the Curve Explained (video)

In this scenario, the goal is to create an ML model to predict which newly uploaded videos will be the most popular on a video sharing website. The result that should be used to determine whether the model is successful is the one that best aligns with the business objective and the evaluation metric. Option C is the correct answer because it defines the most popular videos as the ones that have the highest watch time within 30 days of being uploaded, and it sets a high accuracy threshold of 95% for the model prediction.

Option C: The model predicts 95% of the most popular videos measured by watch time within 30 days of being uploaded. This option is the best result for the scenario because it reflects the business objective and the evaluation metric. The business objective is to prioritize the videos that will attract and retain the most viewers on the website. The watch time is a good indicator of the viewer engagement and satisfaction, as it measures how long the viewers watch the videos. The 30-day window is a reasonable time frame to capture the popularity trend of the videos, as it accounts for the initial interest and the viral potential of the videos. The 95% accuracy threshold is a high standard for the model prediction, as it means that the model can correctly identify 95 out of 100 of the most popular videos based on the watch time metric.

Option A: The model predicts videos as popular if the user who uploads them has over 10,000 likes. This option is not a good result for the scenario because it does not reflect the business objective or the evaluation metric. The business objective is to prioritize the videos that will be the most popular on the website, not the users who upload them. The number of likes that a user has is not a good indicator of the popularity of their videos, as it does not measure the viewer engagement or satisfaction with the videos. Moreover, this option does not specify a time frame or an accuracy threshold for the model prediction, making it vague and unreliable.

Option B: The model predicts 97.5% of the most popular clickbait videos measured by number of clicks. This option is not a good result for the scenario because it does not reflect the business objective or the evaluation metric. The business objective is to prioritize the videos that will be the most popular on the website, not the videos that have the most misleading or sensational titles or thumbnails. The number of clicks that a video has is not a good indicator of the popularity of the video, as it does not measure the viewer engagement or satisfaction with the video content. Moreover, this option only focuses on the clickbait videos, which may not represent the majority or the diversity of the videos on the website.

Option D: The Pearson correlation coefficient between the log-transformed number of views after 7 days and 30 days after publication is equal to 0. This option is not a good result for the scenario because it does not reflect the business objective or the evaluation metric. The business objective is to prioritize the videos that will be the most popular on the website, not the videos that have the most consistent or inconsistent number of views over time. The Pearson correlation coefficient is a metric that measures the linear relationship between two variables, not the popularity of the videos. A correlation coefficient of 0 means that there is no linear relationship between the log-transformed number of views after 7 days and 30 days, which does not indicate whether the videos are popular or not. Moreover, this option does not specify a threshold or a target value for the correlation coefficient, making it meaningless and irrelevant.

Representation transformation (normalization) is a technique that transforms the features to be on a similar scale, such as between 0 and 1, or with mean 0 and standard deviation 1. This technique can improve the performance and training stability of the neural network model, as it can prevent the gradient optimization from being dominated by features with larger scales, and help the model converge faster and better. There are different types of normalization techniques, such as min-max scaling, z-score scaling, log scaling, etc. You can learn more about normalization techniques from the following web search results:

Normalization | Machine Learning | Google for Developers

NORMALIZATION TECHNIQUES IN TRAINING DNNS: METHODOLOGY, ANALYSIS AND ...

Visualizing Different Normalization Techniques | by Dibya ... - Medium

Data Normalization Techniques: Easy to Advanced (& the Best)

Oversampling is a technique for dealing with imbalanced datasets, where the majority class dominates the minority class. It balances the distribution of classes by increasing the number of samples in the minority class. Oversampling can improve the performance of a classifier by reducing the bias towards the majority class and increasing the sensitivity to the minority class.

In this case, the dataset includes transactions, of which 1% are identified as fraudulent. This means that the fraudulent transactions are the minority class and the non-fraudulent transactions are the majority class. A random forest model trained on this dataset might have a low recall for the fraudulent transactions, meaning that it might miss many of them and fail to detect fraud. This could have a high cost for the bank and its customers.

One way to overcome this problem is to oversample the fraudulent transactions 10 times, meaning that each fraudulent transaction is duplicated 10 times in the training dataset. This would increase the proportion of fraudulent transactions from 1% to about 10%, making the dataset more balanced. This would also make the random forest model more aware of the patterns and features that distinguish fraudulent transactions from non-fraudulent ones, and thus improve its accuracy and recall for the minority class.

For more information about oversampling and other techniques for imbalanced data, see the following references:

Random Oversampling and Undersampling for Imbalanced Classification

Exploring Oversampling Techniques for Imbalanced Datasets

Using bfloat16 instead of float32 can reduce the memory usage and training time of the model, while having minimal impact on the accuracy. Bfloat16 is a 16-bit floating-point format that preserves the range of 32-bit floating-point numbers, but reduces the precision from 24 bits to 8 bits. This means that bfloat16 can store the same magnitude of numbers as float32, but with less detail. Bfloat16 is supported by TPUs and some GPUs, and can be used as a drop-in replacement for float32 in most cases. Bfloat16 can also improve the numerical stability of the model, as it reduces the risk of overflow and underflow errors.

Reducing the global batch size, the number of layers, or the dimensions of the images can also reduce the memory usage and training time of the model, but they can also affect the model's accuracy and performance. Reducing the global batch size can make the model less stable and converge slower, as it reduces the amount of information available for each gradient update. Reducing the number of layers can make the model less expressive and powerful, as it reduces the depth and complexity of the network. Reducing the dimensions of the images can make the model less accurate and robust, as it reduces the resolution and quality of the input data.Reference:

Bfloat16: The secret to high performance on Cloud TPUs

Bfloat16 floating-point format

How does Batch Size impact your model learning