You are viewing documentation for an older version.

Click here to view the latest documentation.

Predictions (Early Access)

Predictions is part of our Early Access Program, where we work with early users and customers to test new features and get feedback before making them generally available. These features are functional but can change as we improve them. We recommend connecting with our team before running them in production.

Contact us to get access to this feature.

Predictions extends Profiles’ standard feature development functionality and lets you easily create predictive features in your warehouse. You can predict features like:

• Is a customer likely to churn in the next 30 days?
• Will a user make a purchase in the next 7 days?
• Is a lead going to convert?
• How much is a user likely to spend in the next 90 days?

Finally, you can add the predicted feature to user profiles in your warehouse automatically and deliver ML-based segments and audiences to your marketing, product, and customer success teams.

Use cases

This section covers some common Predictions use cases.

Churn prediction

Predicting churn is one of the crucial initiatives across businesses. Without a predicted churn score, your actions are reactive, whereas you can act proactively with a user trait like is_likely_to_churn. Once you have such features, you can activate them with the appropriate outreach programs to prevent user churn.

Customer LTV prediction

Predictions helps you understand your customers’ purchasing behavior over time. You can predict how much amount a particular customer is likely to spend within the prediction time range.

Prerequisites

• You must be using a Snowflake or Redshift warehouse.
• You must set up a standard Profiles project with a feature table model.
• Optional: If you are using Snowflake, you might need to create a Snowpark-optimized warehouse if your dataset is significantly large.

Project setup

This section highlights the project setup steps for the churn prediction and LTV models.

Churn prediction

Setting up Predictions for predicting churn involves four easy-to-follow steps:

1. Set up a feature table with labels
2. Configure training parameters to generate the predictive features.
3. Configure prediction parameters to generate the predictive features.
4. Schedule periodic Predictions to generate the predictive features.

1. Set up a feature_table_model

Follow the Feature table guide to start with a basic Profiles project. The feature you want to predict should be a part of the feature table in the project.

For example, to predict 30-day inactive churn in advance, you should define it in the feature table so that the model knows how to compute this for historic users.

entity_var:
  name: churn_30_days
  select: case when days_since_last_seen >= 30 then 1 else 0 end

2. Training

RudderStack simplifies your training configuration to a set of parameters. Start with a python_model type and mention the following parameters:

train:
    file_extension: .json
    file_validity: 168h
    inputs: &inputs
      - packages/feature_table/models/shopify_user_features
    config:
      data: &model_data_input_configs
        label_column: churn_30_days
        label_value: 1
        prediction_horizon_days: 30
        output_profiles_ml_model: *model_name
        eligible_users: ''
        inputs: *inputs
        entity_column: user_main_id
        recall_to_precision_importance: 1.0
      preprocessing: 
        ignore_features: [name, gender, device_type]

3. Prediction

In your python_model, mention the following parameters:

predict:
    inputs:
      - packages/feature_table/models/shopify_user_features
    config:
      data: *model_data_input_configs
      outputs:
        column_names:
          percentile: &percentile_name percentile_churn_score_7_days
          score: churn_score_7_days
        feature_meta_data: &feature_meta_data
          features:
            - name: *percentile_name
              description: 'Percentile of churn score. Higher the percentile, higher the probability of churn'

4. Scheduling

Contact us to enable this feature for your account.

1. Upload your project to a GitHub repository.
2. Create a Profiles project in the RudderStack dashboard. Use the GitHub repository to set up the project.
3. Schedule your project with the required cadence. Note that this schedule is for prediction.

Trainings are scheduled as per your configuration of the file_validity parameter in the training section of your project.

LTV models

While the default labels in the churn prediction model are Boolean, Profiles also lets you predict a continuous variable like revenue or LTV. The configuration is almost similar to churn prediction with some minor adjustments.

Set up a feature_table_model

The steps are similar to the setup for churn prediction.

Training

Start with a python_model type and specify the following parameters:

train:
    file_extension: .json
    file_validity: 168h
    inputs: &inputs
      - packages/feature_table/models/shopify_user_features
    config:
      data: &model_data_input_configs
        label_column: amount_spent_past_7_days
        task: regression
        prediction_horizon_days: 7
        output_profiles_ml_model: *model_name
        eligible_users: ''
        inputs: *inputs
        entity_column: user_main_id
      preprocessing: 
        ignore_features: [name, gender, device_type]

Note that:

• Unlike churn prediction, you should not specify the label_value and recall_to_precision_importance labels.
• The LTV model introduces a new parameter called task which you must set to regression. Profiles assumes a classification model by default, unless explicitly specified otherwise.

Prediction

The python_model for the LTV use case remains almost the same as churn prediction, except some minor changes:

predict:
    inputs:
      - packages/feature_table/models/shopify_user_features
    config:
      data: *model_data_configs
      preprocessing: *model_prep_configs
      outputs:
        column_names:
          percentile: &percentile_name percentile_predicted_amount_spent
          score: predicted_amount_spent
        feature_meta_data: &feature_meta_data
          features:
            - name: *percentile_name
              description: 'Percentile of predicted future LTV. Higher the percentile, higher the expected LTV.'

The label column is missing in the above predict block as the score parameter captures the actual prediction value and a Boolean flag is not meaningful in regression use cases.

Scheduling

The steps are similar to the setup for churn prediction.

Results

Final output or the predicted features are pushed to your customer360 table. Use the Explorer tab to check the predicted value for any given user along with the historical values up to last 5 runs.

To check the predicted value for a given user:

1. In the Preview section, go to the Predictive features tab.
2. Check the user profile for which the predictive feature has a value.
3. Search the USER_MAIN_ID of the profile in Profile viewer.

The value of the predictive feature is a probability. You can consider it as true or false based on your threshold.

All your predictive features are listed separately in the Overview tab of your Profiles project. You can check the logs of each run in the artifacts directory (available in the History tab of your Profiles project).

Contact us

Predictions is a part of RudderStack’s Early Access Program. Contact us to get access to this feature.

FAQ

Is there a project to understand Predictions further?

You can check the Shopify churn model that builds a churn prediction score on top of the Shopify library project.

Questions? Contact us by email or on Slack