snippets for xai bqml and looker post

README.md

@@ -26,6 +26,7 @@ From sample dataset to activation, these componentized patterns are designed to
 ### Financial use cases
 * Fraud detection
   * How to build a real-time credit card fraud detection solution. ([Code][ccfraud_code] | [Blogpost][ccfraud_techblog] | [Video][ccfraud_video])
+  * How to use explainable AI to understand a fraud prediction. ([Code][xai_code] | [Blogpost][xai_blog])
 
 
 [gaming_propen_code]: gaming/propensity-model/bqml
@@ -48,7 +49,8 @@ From sample dataset to activation, these componentized patterns are designed to
 [ccfraud_code]: https://gitlab.qdatalabs.com/uk-gtm/patterns/cc_fraud_detection/tree/master
 [ccfraud_techblog]: https://cloud.google.com/blog/products/data-analytics/how-to-build-a-fraud-detection-solution
 [ccfraud_video]: https://youtu.be/qQnxq3COr9Q
-
+[xai_code]: finance/explainable-fraud-model-bqml-looker
+[xai_blog]: https://cloud.google.com/blog/products/data-analytics/explainable-ai-using-bigquery-machine-learning-and-looker 
 
 
 

finance/explainable-fraud-model-bqml-looker/create_models.sql

@@ -0,0 +1,11 @@
+CREATE OR REPLACE MODEL retail_banking.fraud_prediction
+    OPTIONS(model_type='logistic_reg', labels=['is_fraud']) AS
+        SELECT * EXCEPT(trans_id)
+        FROM retail_banking.training_data
+        -- Account for class imbalance. Alternatively, use AUTO_CLASS_WEIGHTS=True in the model options
+        WHERE (is_fraud IS TRUE) OR
+        (is_fraud IS NOT TRUE 
+            AND rand() <=(
+                SELECT COUNTIF(is_fraud)/COUNT(*) FROM retail_banking.training_data
+        )
+);

finance/explainable-fraud-model-bqml-looker/create_training_data.sql

@@ -0,0 +1,29 @@
+
+CREATE OR REPLACE TABLE retail_banking.training_data as (
+    SELECT
+        card_transactions.trans_id AS trans_id,
+        card_transactions.is_fraud AS is_fraud,
+        --amount for transaction: higher amounts are more likely to be fraud
+        cast(card_transactions.amount as FLOAT64) AS card_transactions_amount,
+
+        --distance from the customers home: further distances are more likely to be fraud
+        ST_DISTANCE((ST_GEOGPOINT((cast(card_transactions.merchant_lon as FLOAT64)),
+                (cast(card_transactions.merchant_lat as FLOAT64)))), 
+            (ST_GeogPoint((cast(SPLIT(client.address,'|')[
+                OFFSET
+                    (4)] as float64)),
+                (cast(SPLIT(client.address,'|')[
+                    OFFSET
+                    (3)] as float64))))) AS card_transactions_transaction_distance,
+
+        --hour that transaction occured: fraud occurs in middle of night (usually between midnight and 4 am)
+        EXTRACT(HOUR FROM TIMESTAMP(CONCAT(card_transactions.trans_date,' ',card_transactions.trans_time)) ) AS card_transactions_transaction_hour_of_day
+
+    FROM `looker-private-demo.retail_banking.card_transactions` AS card_transactions
+    LEFT JOIN `looker-private-demo.retail_banking.card` AS card 
+        ON card.card_number = card_transactions.cc_number
+    LEFT JOIN `looker-private-demo.retail_banking.disp` AS disp 
+        ON card.disp_id = disp.disp_id
+    LEFT JOIN `looker-private-demo.retail_banking.client`AS client 
+        ON disp.client_id = client.client_id       
+);

finance/explainable-fraud-model-bqml-looker/explain_hypothetical_data.sql

@@ -0,0 +1,21 @@
+SELECT * FROM 
+    ML.EXPLAIN_PREDICT(MODEL retail_banking.fraud_prediction, (
+            SELECT 
+                '001' as trans_id, 
+                500.00 as card_transactions_amount, 
+                600 as card_transactions_transaction_distance, 
+                2 as card_transactions_transaction_hour_of_day
+        UNION ALL
+            SELECT 
+                '002' as trans_id, 
+                5.25 as card_transactions_amount, 
+                2 as card_transactions_transaction_distance, 
+                13 as card_transactions_transaction_hour_of_day
+        UNION ALL
+            SELECT 
+                '003' as trans_id, 
+                175.50 as card_transactions_amount, 
+                45 as card_transactions_transaction_distance,
+                10 as card_transactions_transaction_hour_of_day
+    ), STRUCT(0.55 AS threshold)
+)