使用SQL型Logistic回歸預測客戶流失

CREATE MODEL retention_model_logistic_reg
TRANSFORM (
  vector_assembler(array(total_purchases, total_revenue, avg_order_value, customer_lifetime, days_since_last_purchase, purchase_frequency)) features
  -- Combines selected customer metrics into a feature vector for model training
)
OPTIONS (
  MODEL_TYPE = 'logistic_reg',  -- Specifies logistic regression as the model type
  LABEL = 'churned'             -- Defines the target label for churn classification
)
AS
WITH customer_features AS (
    SELECT
       identityMap['ECID'][0].id AS customer_id,  -- Extract the unique customer ID from identityMap
       AVG(COALESCE(productListItems.priceTotal[0], 0)) AS avg_order_value,  -- Calculates the average order value, and handles null values with COALESCE
       SUM(COALESCE(productListItems.priceTotal[0], 0)) AS total_revenue, -- The sum of all purchase values per customer
       COUNT(COALESCE(productListItems.quantity[0], 0)) AS total_purchases,  -- The total number of items purchased by the customer
       DATEDIFF(MAX(timestamp), MIN(timestamp)) AS customer_lifetime,  -- The days between first and last recorded purchase
       DATEDIFF(CURRENT_DATE, MAX(timestamp)) AS days_since_last_purchase,  -- The days since the last purchase event
       COUNT(DISTINCT CONCAT(YEAR(timestamp), MONTH(timestamp))) AS purchase_frequency  -- The count of unique months with purchases
    FROM
        webevents
    WHERE EXISTS(productListItems, value -> value.priceTotal > 0)  -- Filters transactions with valid total price
      AND commerce.`order`.purchaseID <> ''  -- Ensures the order has a valid purchase ID
    GROUP BY customer_id
),
customer_labels AS (
    SELECT
      identityMap['ECID'][0].id AS customer_id,  -- Extract the unique customer ID for labeling
      CASE
          WHEN DATEDIFF(CURRENT_DATE, MAX(timestamp)) > 90 THEN 1  -- Marks the customer as churned if no purchase occurred in the last 90 days
          ELSE 0
      END AS churned
    FROM
        webevents
    WHERE EXISTS(productListItems, value -> value.priceTotal > 0)
      AND commerce.`order`.purchaseID <> ''
    GROUP BY customer_id
)
SELECT
    f.customer_id,
    f.total_purchases,
    f.total_revenue,
    f.avg_order_value,
    f.customer_lifetime,
    f.days_since_last_purchase,
    f.purchase_frequency,
    l.churned
FROM
    customer_features f
JOIN
    customer_labels l
ON f.customer_id = l.customer_id  -- Join features with churn labels
ORDER BY RANDOM()  -- Shuffles rows randomly for training
LIMIT 500000;  -- Limit the dataset to 500,000 rows for model training

SELECT *
FROM model_evaluate(retention_model_logistic_reg, 1,
WITH customer_features AS (
    SELECT
       identityMap['ECID'][0].id AS customer_id,
       AVG(COALESCE(productListItems.priceTotal[0], 0)) AS avg_order_value,
       SUM(COALESCE(productListItems.priceTotal[0], 0)) AS total_revenue,
       COUNT(COALESCE(productListItems.quantity[0], 0)) AS total_purchases,
       DATEDIFF(MAX(timestamp), MIN(timestamp)) AS customer_lifetime,
       DATEDIFF(CURRENT_DATE, MAX(timestamp)) AS days_since_last_purchase,
       COUNT(DISTINCT CONCAT(YEAR(timestamp), MONTH(timestamp))) AS purchase_frequency
    FROM
        webevents
    WHERE EXISTS(productListItems, value -> value.priceTotal > 0)
      AND commerce.`order`.purchaseID <> ''
    GROUP BY customer_id
),
customer_labels AS (
    SELECT
      identityMap['ECID'][0].id AS customer_id,
      CASE
          WHEN DATEDIFF(CURRENT_DATE, MAX(timestamp)) > 90 THEN 1
          ELSE 0
      END AS churned
    FROM
        webevents
    WHERE EXISTS(productListItems, value -> value.priceTotal > 0)
      AND commerce.`order`.purchaseID <> ''
    GROUP BY customer_id
)
SELECT
    f.customer_id,
    f.total_purchases,
    f.total_revenue,
    f.avg_order_value,
    f.customer_lifetime,
    f.days_since_last_purchase,
    f.purchase_frequency,
    l.churned
FROM
    customer_features f
JOIN
    customer_labels l
ON f.customer_id = l.customer_id); -- Joins customer features with churn labels

SELECT *
FROM model_predict(retention_model_logistic_reg, 1,  -- Applies the trained model for churn prediction
WITH customer_features AS (
    SELECT
       identityMap['ECID'][0].id AS customer_id,
       AVG(COALESCE(productListItems.priceTotal[0], 0)) AS avg_order_value,
       SUM(COALESCE(productListItems.priceTotal[0], 0)) AS total_revenue,
       COUNT(COALESCE(productListItems.quantity[0], 0)) AS total_purchases,
       DATEDIFF(MAX(timestamp), MIN(timestamp)) AS customer_lifetime,
       DATEDIFF(CURRENT_DATE, MAX(timestamp)) AS days_since_last_purchase,
       COUNT(DISTINCT CONCAT(YEAR(timestamp), MONTH(timestamp))) AS purchase_frequency
    FROM
        webevents
    WHERE EXISTS(productListItems, value -> value.priceTotal > 0)  -- Ensures only valid purchase data is considered
      AND commerce.`order`.purchaseID <> ''
    GROUP BY customer_id
),
customer_labels AS (
    SELECT
      identityMap['ECID'][0].id AS customer_id,
      CASE
          WHEN DATEDIFF(CURRENT_DATE, MAX(timestamp)) > 90 THEN 1  -- Identify customers who have not purchased in the last 90 days
          ELSE 0
      END AS churned
    FROM
        webevents
    WHERE EXISTS(productListItems, value -> value.priceTotal > 0)
      AND commerce.`order`.purchaseID <> ''
    GROUP BY customer_id
)
SELECT
    f.customer_id,
    f.total_purchases,
    f.total_revenue,
    f.avg_order_value,
    f.customer_lifetime,
    f.days_since_last_purchase,
    f.purchase_frequency,
    l.churned
FROM
    customer_features f
JOIN
    customer_labels l
ON f.customer_id = l.customer_id);  -- Matches features with their churn labels for prediction