Creating classification and regression predictions

In this article, you will get know how to set up your first prediction based on regression and/or classification model.

  • Regression model is suited for creating predictions that return numerical results. It’s best used in cases such as:
    • Predicting the amount of money spent by particular group of customers in a defined time range.
    • Predicting the amount of items purchased in the defined time range.
  • Classification model is suited for performing the analyses that return the true/false (or 1/0) values. It’s best used when you want to find answers to the questions such as:
    • Will a customer belong to a particular group of customers?
    • Will a customer leave in the next 30 days?

You can create a prediction from scratch or by using predefined scenarios:

  • Find customers who are likely to leave
  • Find customers who are likely to unsubscribe
  • Find customers who are likely to generate the most value or take significant actions
  • Find customers who belong to a class
  • Find customers who will buy through a specific communication channel

Predefined scenarios are configured using a user-friendly wizard that allows you to adjust scenario settings according to your preferences. This is done through a step-by-step configuration form, which provides hints on the interface to guide you. The option of creating a prediction based on a predefined scenario is available in the initial part of creating the prediction (Select the model type).

Prerequisites


  • Enable the Custom predictions. It usually takes several hours to initialize the Predictions module and perform necessary calculations.
  • You should know how to create segmentations or have a segment of customers ready for use as the prediction target.
  • Create an expression based on which the prediction will be made. For example, if you use an expression that calculates Email Open Rate in last 30 days, you will get a prediction for Email Open Rate for the next 30 days.

Configure a prediction


Select the model type

  1. Go to Image presents the Prediction icon > New prediction.
  2. On the pop-up select one of the following options:
    • Use predefined prediction - This option allows you to use the ready-made prediction scenarios and adjust its settings to your preferences in a user-friendly, step-by-step configuration form by following hints shown on the interface. If you select this option, you can skip reading the rest of the article.
    • Create from scratch - This option allows you to create your own configuration of the prediction to achieve your custom business objective. If you use this option, select one of the following models:
      • Regression - to create a prediction that returns a numerical value
      • Classification - to create a prediction that returns a true/false value

Select customers to be analyzed


Select the audience for whom you want to prepare a prediction.

Selecting an audience to be analyzed
Selecting a group of customers
  1. In the Audience section, click Define.
  2. Click Choose segmentation and perform one of the following actions:
    • From the list, select an existing segmentation.
      You can use the search field.
    • To define a new group segmentation, at the bottom of the dropdown list, click the Create new button.
      1. In the Segmentation name field, enter the name of the group of customers and click Next step.
      2. Build the segmentation.
        See this article for instructions.
      3. Click Create segmentation.
  3. Confirm by clicking Apply.

Choose the prediction target


Select the expression based on which the prediction will be made. For example, if you select an expression that calculates Email Open Rate in last 30 days, you will get prediction for Email Open Rate for the next 30 days.

Selecting an expression to define the scope of prediction
Selecting an expression to define the scope of prediction
  1. From the dropdown list, select the expression you prepared previously.
  2. Confirm by clicking Apply.

Select events


Events are customer activities on the website (visits to a website, adding a product to a shopping cart, and so on) and also your activities towards customers (such as sending messages to them). Select the events that the system will use as input to make a prediction. By default, the list already contains the events recommended for the prediction you are creating. The contents of the list is defined while enabling Custom predictions.

Selection of features to support predicting model
Selection of features to support predicting model
  1. To select events, disable the Auto-select events option.
  2. Click Add events.
  3. On the pop-up, select the events.
  4. Confirm your choice by clicking Add.
  5. Confirm the settings in this section by clicking Apply.

Schedule recalculation and result settings


In this section, define the frequency of recalculating the prediction and settings of the event that is generated for customers for whom the prediction is made.

The Settings section
The Settings section
  1. In the Model configuration section, select the number of days after which the prediction is recalculated.

  2. From the How far in advance do you want to make a prediction? dropdown, select the number of days for which you want to make a prediction (calculated from the current date).

  3. To switch on repeatable calculation of the prediction, select the Set up recurring prediction calculation checkbox.
    By default, the calculation is performed once.

  4. In the How would you like to display results? section, select the scale of prediction results - two- or five-point scale.
    The scale is represented as a score_label parameter of the snr.prediction.score event.

  5. Use the slider to define the number of features displayed in the parameters of the event generated for a customer for whom the prediction was made.

    Note: One of the features is a prediction parameter, it is included in the snr.prediction.score event and it can be used while selecting customers for your campaigns.

  6. In the Name field, enter the value of the scoreName parameter of the event generated (snr.prediction.score) when a prediction is made.

    Important:

    You can use the following characters:

    • a-z - only lower case
    • 0-9
    • special characters: . and _

  7. Click Apply.

Save the prediction

To save the prediction:

  • as a draft, click Save.
  • and calculate, click Save & Calculate.
    Tip: You can check the logs of the activated prediction. Learn more.

Result: The output of the prediction is an event generated on the profile card.

Understanding prediction events


The outcomes of predictions based on the regression and classification models are saved as snr.prediction.score events in the Synerise platform. Thanks to that, the predictions events are available in the Analytics, Communication or Automation modules. The schema of a prediction event is standardized. The attributes are explained in the table below the figure.

snr.prediction.score event
snr.prediction.score event
Attribute name Description
action An event type for the prediction - snr.prediction.score
added by An entity that generated this event. In the case of prediction events, it is always Synerise.
modelId A unique identifier of the prediction
modelName The name of the prediction
score_label Prediction output: two- or five-point scale label. You can select the scale in the settings of the prediction.
scoreName (deprecated) The name of the prediction assigned while creating a prediction
score The result of the prediction (based on the expression which was selected as a target while preparing a prediction).
percentile Prediction percentile
classes Available classes: true, false (only for the Classification type of model)
prediction Predicted class (only for Classification type of model)
feature_influence Local feature importance
clientId Customer’s identifier
time Time when a prediction was generated, as a Unix timestamp
configurationVersion Allows to verify whether two consecutive calculations ran with the settings (only for debug)
modelVersion Model version allows to verify whether two consecutive calculations ran on the same model parameters (only for debug)
😕

We are sorry to hear that

Thank you for helping improve out documentation. If you need help or have any questions, please consider contacting support.

😉

Awesome!

Thank you for helping improve out documentation. If you need help or have any questions, please consider contacting support.

Close modal icon Placeholder alt for modal to satisfy link checker