nerdexam
Microsoft

DP-100 · Question #286

Drag and Drop Question You previously deployed a model that was trained using a tabular dataset named training- dataset, which is based on a folder of CSV files. Over time, you have collected the…

The correct answer is training-dataset; predictions-dataset. The data drift monitor requires a baseline dataset, which is the original training data, and a target dataset, which is the newly collected inference data matching the baseline schema for feature comparison.

Train and deploy models

Question

Drag and Drop Question You previously deployed a model that was trained using a tabular dataset named training- dataset, which is based on a folder of CSV files. Over time, you have collected the features and predicted labels generated by the model in a folder containing a CSV file for each month. You have created two tabular datasets based on the folder containing the inference data: one named predictions-dataset with a schema that matches the training data exactly, including the predicted label; and another named features-dataset with a schema containing all of the feature columns and a timestamp column based on the filename, which includes the day, month, and year. You need to create a data drift monitor to identify any changing trends in the feature data since the model was trained. To accomplish this, you must define the required datasets for the data drift monitor. Which datasets should you use to configure the data drift monitor? To answer, drag the appropriate datasets to the correct data drift monitor options. Each source may be used once, more than once, or not at all. You may need to drag the split bar between panes or scroll to view content. NOTE: Each correct selection is worth one point. Answer:

Exhibit

DP-100 question #286 exhibit

Answer Area

Drag items

training-datasetpredictions-datasetfeatures-dataset

Correct arrangement

  • training-dataset
  • predictions-dataset

Explanation

The data drift monitor requires a baseline dataset, which is the original training data, and a target dataset, which is the newly collected inference data matching the baseline schema for feature comparison.

Approach. To configure a data drift monitor, you must define a baseline dataset and a target dataset. The baseline dataset represents the 'known good' state, which is the data the model was originally trained on. In this scenario, 'training-dataset' is explicitly stated as the dataset used to train the model, making it the correct choice for the Baseline dataset. The target dataset is the new data collected over time, which you want to monitor for drift. The question specifies the need to identify changing trends in the 'feature data'. We have two options for inference data: 'predictions-dataset' (features + predicted label, schema matches training data) and 'features-dataset' (features + timestamp). For data drift monitoring, especially when comparing against a training dataset that likely includes both features and a true label, using 'predictions-dataset' is ideal because its schema 'matches the training data exactly, including the predicted label'. This ensures a direct and consistent comparison of all relevant columns, including features, between the baseline and target. Therefore, drag 'training-dataset' to 'Baseline dataset' and 'predictions-dataset' to 'Target dataset'.

Common mistakes.

  • common_mistake. A common mistake would be to use 'features-dataset' for the Target dataset. While 'features-dataset' contains the feature columns, 'predictions-dataset' is explicitly described as having a schema that 'matches the training data exactly, including the predicted label'. When setting up a data drift monitor, having a consistent schema between the baseline and target datasets is crucial for accurate comparison across all features and potentially the label column if it was part of the training data. 'features-dataset' has a timestamp column instead of a label column, making its schema different from the original 'training-dataset', which would complicate a direct, comprehensive drift comparison.

Concept tested. Azure Machine Learning data drift monitoring, specifically understanding the roles of baseline and target datasets for detecting changes in feature distributions over time.

Topics

#Data Drift Monitoring#Model Monitoring#Azure Machine Learning Datasets#Feature Drift

Community Discussion

No community discussion yet for this question.

Sign up to join the discussion
Full DP-100 Practice