Skip to main content

Use Azure Machine Learning for Spam Detection in Email to Case process in Dynamics 365

The following is a guest post from InfoStrat's Dmitri Riz which is also available on the InfoStrat Tech Blog.

Major software platforms like Microsoft Azure are rapidly developing toolkits that simplify access to AI or Machine Learning technologies, making them available to a wider group of developers. This article shows how the new AI Builder tool on Microsoft Azure PowerApps allows for a quick development of a text classifier model and easy integration of the trained model in Dynamics 365 data processing.

I'll use the Dynamics 365 Email to Case process as a use case. Several of our clients are using this process to generate vast amount of customer support cases from incoming emails. Many of spam emails clear regular email filters and end up as cases in Dynamics, forcing Customer Service Reps to manually tag them as spam and close them.

The resulting set of spam cases represents a perfect training set for a binary text classifier model where feature text is represented by the Description field of the case entity and (for this example) the Resolution field contains the tag.

NOTE: I've used the SMS Spam Collection Data Set from UCI machine learning repository to create and classify sample cases using ham and spam tags.

While this post shows how detect SPAM cases, same techniques can be used to build additional text and/or image classifier models to add more detailed data categorization, such as assigning case priority or routing cases to queues or providing automated case responses.
I will set up two components to illustrate this scenario: first I will build and train a text classification model, and then I will build a Flow that will react to the Create event of the Case and will use this model to classify each case as Spam or Ham.

Step 1: Build and Train Text Classification Model with PowerApps AI Builder

Under AI Builder, click on Model, Build a Model and choose 'Text Classification'.

Click on 'Select Text' and follow prompts to select Case entity and Description field as Text data and select Resolution field during Tag selection step. 

Select your language, confirm the data and tag fields are correct and click on Train button to train the model. 

This process may take a while for a large data set.

Wait till the model is trained and confirm the model precision.

The final step to confirm that your new text classifier works is to test it. Click on Quick test button to test it. Try to supply several text values to make sure that the model can resolve all possible tags with sufficiently high confidence score.

Now that we've built our classifier, it's time to move to the second step.

Step 2: Build a Flow to detect SPAM Cases using Text Classification Model

First, we need to create a new Solution. On PowerApps Solutions menu, click +New Solution, enter solution name and save new solution.

Click on +New button and select Flow to create a new Flow.

Select Dynamics 365 connector and 'When a record is created' trigger and select your organization and Case entity.

Click on Add Action button and select Predict action (I renamed it as Classify Case). This action will use our new model to classify the Description text of each newly created case.

Select your model in the dropdown and manually set the Request Payload field to a JSON fragment shown above, referencing Case Description field.

Click on 'Add an action' and select Parse JSON action. It will interpret Predict action's output. Refer to this Microsoft article for the JSON schema that this step requires.

At this point, we should pause and test our solution to see if both components work together. 

Click on Save button and then on Test button. The flow will wait till new Case is created. 

Now, create a new Case record in Dynamics 365 and populate the Description field with a value that should be classified as SPAM. 

After a few moments the Flow run will complete. If everything was set up correctly, you should see Predict step output containing JSON fragment with data show below, where type value should resolve to one of your tags, and the confidence score showing the reliability of the prediction: 

"response": "{\"predictionOutput\":{\"results\":[{\"type\":\"spam\",\"score\":0.9111173}]} 

The last remaining step is to take the prediction tag and apply it to the Case record. 

For this example, I will simply save the tag in Resolution field (see screenshot in the beginning of the post), but you could just as easily cancel the SPAM cases within the same process.

Click on Add action, select Dynamics 365 connection and Update a record action.

Select your Dynamics Environment; select Case entity. 

Select Case (primary key from the newly created case) in Record Identifier field. Scroll down in the list of fields till you see Resolution and map it to type field of Parse JSON action.

Popular posts from this blog

Key Concepts for Microsoft Dynamics 365: Tenant, Instance, App and Solution

Updated 8/15/2022 To understand Microsoft Dynamics 365 (formerly Dynamics CRM) and Power Apps, you need to learn some new terms and concepts that may be a bit different from what you know from databases and solutions that are hosted on premises.  These concepts also apply to Power Apps.  The main difference is that with Power Apps you are not starting with a Microsoft app but more of a blank canvas for your custom apps.  This post introduces some key terms and how these concepts are important for planning your implementation. While Dynamics 365 is still available on premises, it is most commonly deployed on the Microsoft cloud.  This blog post discusses only cloud implementations. Microsoft has multiple clouds such as commercial and government community clouds in several countries. We start with a Microsoft tenant .  A tenant is the account you create in the Microsoft Online Services environment (such as Office 365) when you sign up for a subscription. A tenant contains uni

My Favorite Microsoft Power Apps Bloggers and their Blogs

  by James Townsend Updated 7/5/2022 Microsoft Power Apps is one of my favorite subjects, and I enjoy reading blog posts from members of this thriving technical community.  Here are some of my favorite bloggers and their blogs: The Official Microsoft Power Apps Blog   I have to start with the official Microsoft Power Apps blog.  It has many contributors, largely Microsoft program manager, including frequent posters Denise Moran ,  Greg Lindhorst , Kartik Kanakasabesan , and  Adrian Orth .  This is the place to go for product announcements, updates and technical how-to for a broad range of Power Apps topics.  April Dunnam April Dunnam was formerly focused on SharePoint and now devoting herself to Power Platform.  April offers highly understandable explanations of Power Platform, Dataverse and other top Power Apps topics. She joined Microsoft in late 2019 and has a thriving YouTube channel .  Carl De Souza Power Apps Blog and eBook This is one of the most extensive and best organized blo

Understanding Dynamics 365 and Office 365 Admin Roles

Managing Dynamics 365 instances If you run Microsoft Dynamics 365 (formerly Dynamics CRM) in the Microsoft cloud, you need to understand how your Dynamics instances relate to Office 365 and choose which of your administrators receives which roles and permissions to manage Dynamics 365. In on premises deployments, your network administrator would create and delete user accounts.  The Dynamics 365 admin would then assign permissions to users in Dynamics 365. This post explains three administrator roles: Office 365 Global Administrator Dynamics 365 System Administrator Dynamics 365 Service Administrator You may think that the Dynamics 365 system administrator would have power to do all the actions needed to manage Dynamics 365, but this is not the case. What's different in Microsoft cloud deployments is that licenses and user accounts are managed in Office 365 by an Office 365 Global Administrator.  This role is analogous to a network administrator for an on premises