Training models by using AutoML

Unique identifiers like customer_email should be excluded from training datasets because they don't provide any generalizable patterns for prediction. Each value appears only once, so the model cannot learn meaningful relationships. Option A is incorrect because while AutoML has some automation, including unique identifiers can lead to overfitting. Option B is incorrect because one-hot encoding a unique identifier would create as many columns as rows, which is inefficient and unhelpful. Option D is incorrect because the target should be the churn indicator, not the email address.

Vertex AI Data Labeling Service is the most efficient approach as it provides a managed service with quality control, specialist labelers, and integration with Vertex AI. It supports active learning and can significantly reduce labeling time and costs. Option B is inefficient and doesn't scale well for 10,000 images. Option C is incorrect because AutoML Vision requires labeled data for supervised learning tasks like classification. Option D won't work for custom defect detection as pre-trained models aren't trained on your specific defect types.

Creating derived temporal features like day_of_week, month, is_holiday, and hour_of_day helps the model capture seasonal patterns and cyclical trends that are crucial for forecasting. These features make temporal patterns explicit and easier for the model to learn. Option A is incorrect as timestamps contain valuable information for forecasting. Option C (Unix epoch) doesn't make patterns obvious to the model. Option D is incorrect because aggregating to yearly intervals loses the granularity needed for accurate forecasting and eliminates important seasonal patterns.

AutoML Vision automatically handles image preprocessing including resizing, normalization, and augmentation. You should upload images in their original format and let AutoML handle the optimization. Option A is unnecessary and may reduce image quality. Option B would severely limit your dataset and isn't required. Option D is incorrect because converting to grayscale loses color information that might be important for classification, and AutoML can handle color images efficiently.

Checking for class imbalance is the first step when you notice poor performance on specific classes. If negative sentiments are underrepresented in your training data, the model won't learn to identify them effectively. You may need to collect more negative examples or apply techniques like class weighting. Option A might help but won't solve an imbalance problem. Option C is premature before understanding the root cause. Option D doesn't apply well to text classification where the text itself is the primary input.

Using Cloud DLP to de-identify or tokenize PII before training is the best practice for responsible AI. You should transform PII into aggregate features (e.g., zip code instead of full address, age range instead of birthdate) that preserve predictive value while protecting privacy. Option A violates privacy principles and regulations. Option C still exposes PII during training even if encrypted. Option D is better but storing IDs of sensitive data can still enable re-identification and doesn't fully address the privacy concern.

AutoML Tables automatically performs feature selection and provides feature importance scores in the evaluation results. This helps identify which features contribute most to predictions. You can review these insights and optionally remove low-importance features for model simplification. Option A is incorrect as AutoML handles feature importance internally. Option C is incorrect because AutoML doesn't require manual PCA. Option D is incorrect because AutoML automatically handles feature normalization and scaling.

For time-series forecasting, chronological splitting is essential to prevent data leakage. Training should use older historical data, and validation/testing should use more recent data to simulate real-world prediction scenarios. Option A (random split) causes data leakage as future information would be in the training set. Option B is backwards and would train on recent data to predict the past. Option D (random k-fold) also causes temporal leakage and doesn't respect the time-series nature of the data.

Segmenting long videos into shorter, meaningful clips improves training efficiency and model accuracy. Each clip should represent a single class or interaction type. This helps the model focus on relevant content and reduces processing costs. Option A is inefficient and may confuse the model with mixed content. Option C is unnecessary as AutoML handles frame rate normalization. Option D loses temporal information that's crucial for understanding video content and interaction context.

The large gap between training (99%) and test (65%) accuracy is a classic sign of overfitting, where the model memorizes training data rather than learning generalizable patterns. Solutions include early stopping, regularization, or collecting more diverse training data. In AutoML, you can also try reducing training time or using a simpler model. Option A would likely worsen overfitting. Option C might help evaluation but doesn't address the core overfitting problem. Option D (underfitting) is incorrect because underfitting shows poor performance on both training and test sets.