Churn Model Training and Predictions
Churn Model Training
The Churn Model Training batch trains a customer churn prediction model and stores the trained model for use in future prediction workflows.
Input Parameter
Model Name must be a unique, alphanumeric model name. Spaces and special characters are not supported. Maximum length: 20 characters.
Source Data
The training dataset is sourced from VW_ML_CHURN_MODEL_DATASET table.
Prerequisites
- Complete data loading.
- Successfully complete Run Chart execution, including customer segmentation.
- Configure the dataset view to return data for both current and previous segmentation dates in FSI_BI_SEGMENTATION.
- Ensure the dataset has both churn and non-churn cases.
- The process uses Synthetic Minority Over-sampling Technique (SMOTE) to improve class balance, but highly skewed data can affect model quality.
Churn Case Definition
- The customer’s balance is zero.
- All customer accounts have an account-closure date.
- The customer exists in the previous dataset but is absent from the latest dataset.
All other customers are classified as non-churn (negative) cases.
Output
- The trained model is stored in FSI_CS_MODEL_STORE table.
- One record is created for each unique model name.
- Reusing an existing model name updates the stored model and its performance metrics.
- The model is available for downstream prediction processes.
Product and Channel Mappings
Product mappings are configured in Environment Properties:
- Property Class: ML_MODELS
- Property Code: PRODUCT_ML_MAPPING
Channel mappings are configured in Environment Properties:
- Property Class: ML_MODELS
- Property Code: CHANNEL_ML_MAPPING
Note:
These mappings are customer-specific. Refer to the Property Code Details table in the Machine Learning Model Properties section.Churn Predictions
The Churn Predictions batch uses a previously trained churn model to generate customer churn predictions. Results are stored in the database and is available in Churn Model Predictions page.
Prerequisites
- Train a churn model through PA Churn Model Training.
- Confirm the specified model name exists in the model repository.
- Ensure all required source data is loaded and available.
Processing
- Retrieves the trained model for the specified model name.
- Applies the model to the latest customer data.
- Generates churn predictions for eligible customers.
Output
- Results include churn scores generated by the trained model.
- Predictions are stored in FSI_CS_CHURN_PREDICTIONS table.
- Predictions are available for reporting and downstream analysis.
Viewing Predictions
Use Churn Model Predictions interface to view the latest churn prediction results generated by the PA Churn Predictions batch.
Note:
- PA Churn Predictions requires a successfully trained model.
- Predictions cannot be generated if the specified model name does not exist.
- Use the exact same model name for training and prediction.