Merge pull request #50454 from ShawnKupfer/WB1791

JillGrant615 · web-flow · commit b2dce87585f0 · 2025-05-22T08:46:20.000-06:00
AB#1013451: Refine and test machine learning models
diff --git a/learn-pr/azure/test-machine-learning-models/1-introduction.yml b/learn-pr/azure/test-machine-learning-models/1-introduction.yml
@@ -4,7 +4,7 @@ title: Introduction
 metadata:
   title: Introduction
   description: Introduction to the introduction to regression module.
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
diff --git a/learn-pr/azure/test-machine-learning-models/2-normalization-and-standardization.yml b/learn-pr/azure/test-machine-learning-models/2-normalization-and-standardization.yml
@@ -4,7 +4,7 @@ title: Normalization and standardization
 metadata:
   title: Normalization and standardization
   description: Conceptual unit introducing normalization and standardization in machine learning
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
diff --git a/learn-pr/azure/test-machine-learning-models/4-test-training-datasets.yml b/learn-pr/azure/test-machine-learning-models/4-test-training-datasets.yml
@@ -4,7 +4,7 @@ title: Test and training datasets
 metadata:
   title: Test and training datasets
   description: Conceptual unit about testing and training datasets in machine learning
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
diff --git a/learn-pr/azure/test-machine-learning-models/5-exercise-test-training-datasets.yml b/learn-pr/azure/test-machine-learning-models/5-exercise-test-training-datasets.yml
@@ -4,7 +4,7 @@ title: Exercise - Test and train datasets
 metadata:
   title: Exercise - Test and train datasets
   description: Exercise unit testing and training datasets in machine learning
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
diff --git a/learn-pr/azure/test-machine-learning-models/6-nuance-tests.yml b/learn-pr/azure/test-machine-learning-models/6-nuance-tests.yml
@@ -4,7 +4,7 @@ title: Nuances of test sets
 metadata:
   title: Nuances of test sets
   description: Conceptual unit about nuances of test sets in machine learning
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
diff --git a/learn-pr/azure/test-machine-learning-models/7-exercise-test-set-nuances.yml b/learn-pr/azure/test-machine-learning-models/7-exercise-test-set-nuances.yml
@@ -1,10 +1,10 @@
 ### YamlMime:ModuleUnit
 uid: learn.machinelearning.test-machine-learning-models.exercise-test-set-nuances
-title: Exercise – Test set nuances
+title: Exercise - Test set nuances
 metadata:
-  title: Exercise – Test set nuances
+  title: Exercise - Test set nuances
   description: Exercise unit about test set nuances in machine learning
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
diff --git a/learn-pr/azure/test-machine-learning-models/8-knowledge-check.yml b/learn-pr/azure/test-machine-learning-models/8-knowledge-check.yml
@@ -4,7 +4,7 @@ title: Module assessment
 metadata:
   title: Module assessment
   description: Multiple-choice questions
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
@@ -28,7 +28,7 @@ quiz:
     - content: "Underfitting has occurred, and your model isn't accurate enough. You should keep training."
       isCorrect: false
       explanation: "Incorrect. Continuing to train your model when you already have good performance on your training set won't improve your performance. You need to find ways to improve performance on your test set."
-    - content: "Overfitting has occurred, and your model isn't performing well on new data outside training. You could stop training earlier, or gather more diverse data."
+    - content: "Overfitting has occurred, and your model isn't performing well on new data outside training. You could stop training earlier or gather more diverse data."
       isCorrect: true
       explanation: "Correct. Overfitting has likely occurred, and you can adjust your training to improve performance on your test set. You should consider if you need more diverse training data, or if you're training for too long."
     - content: "Your model is fine. You need to use your training data to test your model instead."
diff --git a/learn-pr/azure/test-machine-learning-models/9-summary.yml b/learn-pr/azure/test-machine-learning-models/9-summary.yml
@@ -4,7 +4,7 @@ title: Summary
 metadata:
   title: Summary
   description: An overview of the content covered in the module.
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: unit
diff --git a/learn-pr/azure/test-machine-learning-models/includes/1-introduction.md b/learn-pr/azure/test-machine-learning-models/includes/1-introduction.md
@@ -2,9 +2,9 @@ The way we train models is by no means a perfectly automated process. Training's
 
 ## Scenario: Training avalanche rescue dogs
 
-Throughout this module, we’ll be using the following example scenario to explain underfitting and overfitting. This scenario is designed to provide an example for how you might meet these concepts while programming for yourself. Keep in mind that these principles generally apply to almost all types of models, not just those we work with here.
+Throughout this module, we'll be using the following example scenario to explain underfitting and overfitting. This scenario is designed to provide an example for how you might meet these concepts while programming for yourself. Keep in mind that these principles generally apply to almost all types of models, not just those we work with here.
 
-It’s time for your charity to train a new generation of dogs in how to find hikers swept up by avalanches. There's debate in the office as to which dogs are best; is a large dog better than a smaller dog? Should the dogs be trained when they're young or when they're more mature? Thankfully, you have statistics on rescues performed over the last few years that you can look to. Training dogs is expensive, though, and you need to be sure that your dog-picking criteria are sound.
+It's time for your charity to train a new generation of dogs in how to find hikers swept up by avalanches. There's debate in the office as to which dogs are best; is a large dog better than a smaller dog? Should the dogs be trained when they're young or when they're more mature? Thankfully, you have statistics on rescues performed over the last few years that you can look to. Training dogs is expensive, though, and you need to be sure that your dog-picking criteria are sound.
 
 ## Prerequisites
 
diff --git a/learn-pr/azure/test-machine-learning-models/includes/2-normalization-and-standardization.md b/learn-pr/azure/test-machine-learning-models/includes/2-normalization-and-standardization.md
@@ -2,23 +2,23 @@ _Feature Scaling_ is a technique that changes the range of values that a feature
 
 ## Normalization versus standardization
 
-_Normalization_ means to scale values so that they all fit within a certain range, typically 0–1. For example, if you had a list of people’s ages that were 0, 50, and 100 years, you could normalize by dividing the ages by 100, so that your values were 0, 0.5, and 1.
+_Normalization_ means to scale values so that they all fit within a certain range, typically 0–1. For example, if you had a list of people's ages that were 0, 50, and 100 years, you could normalize by dividing the ages by 100 so that your values were 0, 0.5, and 1.
 
-_Standardization_ is similar, but instead, we subtract the mean (also known as the average) of the values and divide by the standard deviation. If you’re not familiar with standard deviation, not to worry, this means that after standardization, our mean value is zero, and about 95% of values fall between -2 and 2.
+_Standardization_ is similar, but instead, we subtract the mean (also known as the average) of the values and divide by the standard deviation. If you're not familiar with standard deviation, not to worry; this means that after standardization, our mean value is zero, and about 95% of values fall between -2 and 2.
 
-There are other ways to scale data, but the nuances of these are beyond what we need to know right now. Let’s explore why we apply _normalization_ or _standardization_.
+There are other ways to scale data, but the nuances of these are beyond what we need to know right now. Let's explore why we apply _normalization_ or _standardization_.
 
 ## Why do we need to scale?
 
-There are many reasons we normalize or standardize data before training. You can understand these more easily with an example. Let’s say we want to train a model to predict whether a dog will be successful at working in the snow. Our data are shown in the following graph as dots, and the trend line we're trying to find is shown as a solid line:
+There are many reasons we normalize or standardize data before training. You can understand these more easily with an example. Let's say we want to train a model to predict whether a dog will be successful at working in the snow. Our data are shown in the following graph as dots, and the trend line we're trying to find is shown as a solid line:
 
 ![Diagram showing scaling in a graph of dog height and rescues starting at 50.](../media/2-normalization-graph.png)
 
 ### Scaling gives learning a better starting point
 
-The optimal line in the preceding graph has two parameters: the intercept, which is 50, the line at x=0, and slope, which is 0.01; each 1000 millimeters increases rescues by 10. Let’s assume we start training with initial estimates of 0 for both of these parameters.
+The optimal line in the preceding graph has two parameters: the intercept, which is 50, the line at x=0, and slope, which is 0.01; each 1000 millimeters increases rescues by 10. Let's assume we start training with initial estimates of 0 for both of these parameters.
 
-If our training iterations are altering parameters by around 0.01 per iteration on average, it takes at least 5000 iterations before the intercept is found: 50 / 0.01 = 5000 iterations. Standardization can bring this optimal intercept is closer to zero, which means we can find it much faster. For example, if we subtract the mean from our label—annual rescues—and our feature—height—the intercept is -0.5, not 50, which we can find about 100 times faster.
+If our training iterations are altering parameters by around 0.01 per iteration on average, it takes at least 5000 iterations before the intercept is found: 50 / 0.01 = 5000 iterations. Standardization can bring this optimal intercept is closer to zero, which means we can find it much faster. For example, if we subtract the mean from our label (annual rescues) and our feature (height) the intercept is -0.5, not 50, which we can find about 100 times faster.
 
 ![Diagram showing scaling in a graph of dog height and rescues starting at 0.](../media/2-normalization-graph-2.png)
 
@@ -42,6 +42,6 @@ When we work with multiple features, having these on a different scale can cause
 
 ## Do I always need to scale?
 
-We don’t always need to scale. Some kinds of models, including the preceding models with straight lines, can be fit without an iterative procedure like gradient descent, so they don't mind features being the wrong size. Other models do need scaling to train well, but their libraries often perform feature scaling automatically.
+We don't always need to scale. Some kinds of models, including the preceding models with straight lines, can be fit without an iterative procedure like gradient descent so they don't mind features being the wrong size. Other models do need scaling to train well, but their libraries often perform feature scaling automatically.
 
 Generally speaking, the only real downsides to normalization or standardization are that it can make it harder to interpret our models and that we have to write slightly more code. For this reason, feature scaling is a standard part of creating machine learning models.
diff --git a/learn-pr/azure/test-machine-learning-models/includes/4-test-training-datasets.md b/learn-pr/azure/test-machine-learning-models/includes/4-test-training-datasets.md
@@ -1,4 +1,4 @@
-The data we use to train a model is often called a _training dataset_. We’ve already seen this in action. Frustratingly, when we use the model in the real world, after training we don’t know for certain how well our model will work. This uncertainty is because it’s possible that our training dataset is different to data in the real world.
+The data we use to train a model is often called a _training dataset_. We've already seen this in action. Frustratingly, when we use the model in the real world, we don't know for certain how well our model will work after training. This uncertainty is because it's possible that our training dataset is different to data in the real world.
 
 ## What is overfitting?
 
@@ -10,7 +10,7 @@ We can avoid overfitting several ways. The simplest way is to have a simpler mod
 
 ![Diagram showing a plot graph of dog height and rescues.](../media/4-overfitting-graph.png)
 
-Let’s say we collect information about only five dogs, though, and use that as our training dataset to fit a complex line. If we can do so, we can fit it very well:
+Let's say we collect information about only five dogs, though, and use that as our training dataset to fit a complex line. If we can do so, we can fit it very well:
 
 ![Diagram showing a complex line graph using only five dogs height and rescue information.](../media/4-overfitting-graph-2.png)
 
@@ -28,9 +28,9 @@ A complimentary way we can avoid overfitting is to stop training after the model
 
 A test dataset, also called a validation dataset, is a set of data similar to the training dataset. In fact, test datasets are usually created by taking a large dataset and splitting it. One portion is called the training dataset, and the other is called the test dataset.
 
-The job of the training dataset is to train the model; we’ve seen training already. The job of the test dataset is to check how well the model works; it doesn't contribute to training directly.
+The training dataset's job is to train the model; we've seen training already. The test dataset's job is to check how well the model works; it doesn't contribute to training directly.
 
-### OK, but what’s the point?
+### OK, but what's the point?
 
 The point of a test dataset is twofold.
 
diff --git a/learn-pr/azure/test-machine-learning-models/includes/6-nuance-tests.md b/learn-pr/azure/test-machine-learning-models/includes/6-nuance-tests.md
@@ -1,22 +1,22 @@
-Test sets are considered best practice for most aspects of machine learning, though the field is still relatively young, and so exactly how and when is often debated. Let’s go through some things to consider.
+Test sets are considered best practice for most aspects of machine learning, though the field is still relatively young, and so exactly how and when is often debated. Let's go through some things to consider.
 
 ## Test sets can be misleading
 
 Although test sets are helpful to identify overtraining, they can provide us with false confidence. Specifically, test sets are only useful if they reflect data that we expect to see in the real world. For example, our test set is very small, so it won't be representative of the variety of data that we're likely to see in the real world. Test datasets are also only as good as their source. If our test dataset comes from a biased source, our metrics won't reflect how things will behave in the real world.
 
-For example, let’s say we're trying to find the relationship between number of rescues and the age a dog started training. If our test set was only three dogs, it's possible that these dogs aren't a good representation of the wide variety of working dogs in the real world. Also, imagine that we obtained our test set from a single breeder who doesn't know how to work with puppies. Our model might predict that older dogs are best to train, and our test dataset would confirm this, when in fact other trainers might have enormous success with younger animals.
+For example, let's say we're trying to find the relationship between number of rescues and the age a dog started training. If our test set was only three dogs, it's possible that these dogs aren't a good representation of the wide variety of working dogs in the real world. Also, imagine that we obtained our test set from a single breeder who doesn't know how to work with puppies. Our model might predict that older dogs are best to train, and our test dataset would confirm this, when in fact other trainers might have enormous success with younger animals.
 
 ## Test sets aren't free
 
-We’ve already seen that the more training data we have, the less likely our model will overfit. Similarly, the larger the test sets, the more we feel we can trust our test results. However, we usually work with finite amounts of data, and a datapoint can't be in both the training and the test set. This means that as we get larger test sets, we get smaller training datasets and vice versa. Exactly how much data should be sacrificed to appear in the test dataset depends on individual circumstances, with anything between 10-50% being relatively common, depending on the volume of data available.
+We've already seen that the more training data we have, the less likely our model will overfit. Similarly, the larger the test sets, the more we feel we can trust our test results. However, we usually work with finite amounts of data, and a datapoint can't be in both the training and the test set. This means that as we get larger test sets, we get smaller training datasets and vice versa. Exactly how much data should be sacrificed to appear in the test dataset depends on individual circumstances, with anything between 10-50% being relatively common, depending on the volume of data available.
 
 ## Train and test isn't the only approach
 
-It’s worth keeping in mind that train-and-test is common, but not the only widely used approach. Two of the more common alternatives are the *hold-out approach* and *statistical approach* methods.
+It's worth keeping in mind that train-and-test is common, but not the only widely used approach. Two of the more common alternatives are the *hold-out approach* and *statistical approach* methods.
 
 ### The hold-out approach
 
-The hold-out approach is like train-and-test, but instead of splitting a dataset into two, it's split into three: _training_, _test_ (also known as _validation_), and _hold-out._ The training and test datasets are as we’ve described previously. The hold-out dataset is a kind of test set that's used only once, when we're ready to deploy our model for real-world use. In other words, it's not used until we've finished experimenting with different kinds of training regimens, different kinds of models, and so on.
+The hold-out approach is like train-and-test, but instead of splitting a dataset into two, it's split into three: _training_, _test_ (also known as _validation_), and _hold-out._ The training and test datasets are as we've described previously. The hold-out dataset is a kind of test set that's used only once, when we're ready to deploy our model for real-world use. In other words, it's not used until we've finished experimenting with different kinds of training regimens, different kinds of models, and so on.
 
 This approach tackles the fact that we usually experiment with different models and training regimens. For example, we fit a model, find it doesn't work well with the test dataset, change some aspects of the model being trained, and try again until we get a good result. This means we're purposefully altering our model to work for a particular set of data, just like normal training does with the training dataset. By doing this, we can end up with a model that's essentially too overtrained to work on our test dataset.
 
diff --git a/learn-pr/azure/test-machine-learning-models/includes/9-summary.md b/learn-pr/azure/test-machine-learning-models/includes/9-summary.md
@@ -11,4 +11,4 @@ Now that you've reviewed this module, you should be able to:
 [!include[](../../../includes/open-link-in-new-tab-note.md)]
 
 * [ML.NET Tutorial](https://dotnet.microsoft.com/learn/ml-dotnet/get-started-tutorial/intro)
-* [Azure Machine Learning](https://azure.microsoft.com/services/machine-learning/)
+* [Azure Machine Learning](https://azure.microsoft.com/products/machine-learning/)
diff --git a/learn-pr/azure/test-machine-learning-models/index.yml b/learn-pr/azure/test-machine-learning-models/index.yml
@@ -3,18 +3,19 @@ uid: learn.machinelearning.test-machine-learning-models
 metadata:
   title: Refine and test machine learning models
   description: When we think of machine learning, we often focus on the training process. A small amount of preparation before this process can not only speed up and improve learning, but also give us some confidence about how well our models will work when faced with data we have never seen before.
-  ms.date: 05/25/2021
+  ms.date: 05/15/2025
   author: s-polly
   ms.author: scottpolly
   ms.topic: module
+  ms.service: azure-machine-learning
 title: Refine and test machine learning models
 summary: When we think of machine learning, we often focus on the training process. A small amount of preparation before this process can not only speed up and improve learning, but also give us some confidence about how well our models will work when faced with data we have never seen before.
 abstract: | 
   In this module, you will:
   - Define feature scaling.
   - Create and work with test datasets.
   - Articulate how testing models can both improve and harm training.
-prerequisites: Familiarity with machine learning models
+prerequisites: Familiarity with machine-learning models
 iconUrl: /training/achievements/machine-learning/test-machine-learning-models.svg
 levels: 
 - beginner
diff --git a/learn-pr/azure/test-machine-learning-models/notebooks/5-3-exercise-feature-normalization.ipynb b/learn-pr/azure/test-machine-learning-models/notebooks/5-3-exercise-feature-normalization.ipynb
diff --git a/learn-pr/azure/test-machine-learning-models/notebooks/5-5-exercise-test-training-datasets.ipynb b/learn-pr/azure/test-machine-learning-models/notebooks/5-5-exercise-test-training-datasets.ipynb
diff --git a/learn-pr/azure/test-machine-learning-models/notebooks/5-7-exercise-test-set-nuances.ipynb b/learn-pr/azure/test-machine-learning-models/notebooks/5-7-exercise-test-set-nuances.ipynb
