Deploying the Churn Prediction Model in Production

Overview

Deploying a churn prediction model into production is the final step in leveraging machine learning insights to drive business decisions. BigQuery ML simplifies this process by enabling both batch and real-time predictions using SQL queries. This article outlines the deployment steps, compares batch and real-time prediction approaches, and discusses cost management strategies to ensure efficient operations.

---

Batch Predictions vs. Real-Time Predictions

Batch Predictions

Batch predictions are suitable for periodic updates where predictions don’t need to be generated instantly. This approach is ideal for churn prediction in scenarios like:

• Weekly or monthly churn reports.
• Customer retention campaigns.

Advantages:

• Lower costs as predictions are run less frequently.
• Simpler implementation, as they rely on scheduled queries.

Disadvantages:

• Not suitable for use cases requiring immediate insights.

Real-Time Predictions

Real-time predictions are triggered for specific events, such as when a customer interacts with a service. This approach is ideal for:

• Dynamic churn prediction during customer interactions.
• Triggering personalized offers or interventions.

Advantages:

• Instant insights allow for immediate action.
• Enhances customer experience with timely responses.

Disadvantages:

• Higher costs due to constant computation.
• Requires more complex infrastructure, such as event-driven architectures.

---

Prediction Queries in BigQuery

Batch Prediction Query

The following query predicts churn for all customers and writes the results to a BigQuery table for downstream analysis:

CREATE OR REPLACE TABLE `<DATASET_NAME>.predictions` AS
SELECT
  customerID,
  predicted_label AS churn_prediction,
  predicted_probability AS churn_probability,
  *
FROM
  ML.PREDICT(MODEL `<DATASET_NAME>.<MODEL_NAME>`, 
             TABLE `<DATASET_NAME>.<INPUT_TABLE_NAME>`);

Explanation:

• ML.PREDICT: Generates predictions using the specified model.
• predicted_label: Indicates whether a customer is predicted to churn.
• predicted_probability: Provides the confidence score for the prediction.

---

Real-Time Prediction Query

For real-time predictions, a single customer’s data can be used as input:

SELECT
  customerID,
  predicted_label AS churn_prediction,
  predicted_probability AS churn_probability
FROM
  ML.PREDICT(MODEL `<DATASET_NAME>.<MODEL_NAME>`, 
             (SELECT * FROM `<DATASET_NAME>.<INPUT_TABLE_NAME>` WHERE customerID = 'CUST123'));

Explanation:

• Filters the input table to focus on a specific customer (CUST123).
• Returns immediate churn predictions for personalized actions.

---

Integration with Cloud Functions

To enable real-time predictions, you can use Cloud Functions and Pub/Sub to automate the process. For example, trigger a Cloud Function whenever a customer event is logged.

Cloud Function Code Example

import os
from google.cloud import bigquery

def predict_churn(event, context):
    client = bigquery.Client()
    dataset_name = os.environ.get("DATASET_NAME")
    model_name = os.environ.get("MODEL_NAME")
    customer_id = event['attributes']['customerID']

    query = f"""
    SELECT
        predicted_label AS churn_prediction,
        predicted_probability AS churn_probability
    FROM
        ML.PREDICT(MODEL `{dataset_name}.{model_name}`, 
                   (SELECT * FROM `{dataset_name}.customers` WHERE customerID = '{customer_id}'))
    """
    
    results = client.query(query).result()
    for row in results:
        print(f"CustomerID: {customer_id}, Churn Prediction: {row.churn_prediction}, Probability: {row.churn_probability}")

Explanation:

• Event-Driven Architecture: The function is triggered by a Pub/Sub event containing the customerID.
• BigQuery Integration: Executes a query to predict churn for the specified customer.

---

Cost Management

Running predictions at scale can incur significant costs. Here are some strategies to optimize spending:

• Batch Predictions: Schedule predictions during non-peak hours to leverage GCP’s cost-saving features.
• Selective Real-Time Predictions: Limit real-time predictions to high-value customers or critical interactions.
• BigQuery Slot Reservations: For frequent predictions, consider purchasing BigQuery slots to reduce on-demand query costs.

---

Code Snippets for Batch and Real-Time Predictions

Batch Prediction Scheduling with Cloud Scheduler

gcloud scheduler jobs create query churn-batch-predictions \
--schedule="0 2 * * *" \
--statement="CREATE OR REPLACE TABLE `<DATASET_NAME>.predictions` AS SELECT * FROM ML.PREDICT(MODEL `<DATASET_NAME>.<MODEL_NAME>`, TABLE `<DATASET_NAME>.<INPUT_TABLE_NAME>`);" \
--time-zone="UTC" \
--bigquery-use-default-project

Explanation:

• Runs the batch prediction query nightly at 2 AM UTC.
• Stores results in the predictions table for analysis.

---

Deploying a churn prediction model involves careful consideration of use cases, cost management, and infrastructure complexity. By leveraging BigQuery ML’s integration with Cloud Functions and Pub/Sub, you can implement both batch and real-time predictions tailored to business needs.