USE sales
GO

-- Create view used for ML services training stored procedure
CREATE OR ALTER VIEW [dbo].[web_clickstreams_hdfs_book_clicks]
AS
	SELECT
        /* There is bug in TPCx-BB data generator which results in data where all users have purchased books.
        This will not work for the ML training purposes. So we will treat users with 1-5 clicks in the book category as
        not interested in books. */
	  CASE WHEN q.clicks_in_category < 6 THEN 0 ELSE q.clicks_in_category END AS clicks_in_category,
	  CASE WHEN cd.cd_education_status IN ('Advanced Degree', 'College', '4 yr Degree', '2 yr Degree') THEN 1 ELSE 0 END AS college_education,
	  CASE WHEN cd.cd_gender = 'M' THEN 1 ELSE 0 END AS male,
	  COALESCE(cd.cd_credit_rating, 'Unknown') as cd_credit_rating,
	  q.clicks_in_1,
	  q.clicks_in_2,
	  q.clicks_in_3,
	  q.clicks_in_4,
	  q.clicks_in_5,
	  q.clicks_in_6,
	  q.clicks_in_7,
	  q.clicks_in_8,
	  q.clicks_in_9,
	  q.wcs_user_sk
	FROM( 
	  SELECT 
		w.wcs_user_sk,
		SUM( CASE WHEN i.i_category = 'Books' THEN 1 ELSE 0 END) AS clicks_in_category,
		SUM( CASE WHEN i.i_category_id = 1 THEN 1 ELSE 0 END) AS clicks_in_1,
		SUM( CASE WHEN i.i_category_id = 2 THEN 1 ELSE 0 END) AS clicks_in_2,
		SUM( CASE WHEN i.i_category_id = 3 THEN 1 ELSE 0 END) AS clicks_in_3,
		SUM( CASE WHEN i.i_category_id = 4 THEN 1 ELSE 0 END) AS clicks_in_4,
		SUM( CASE WHEN i.i_category_id = 5 THEN 1 ELSE 0 END) AS clicks_in_5,
		SUM( CASE WHEN i.i_category_id = 6 THEN 1 ELSE 0 END) AS clicks_in_6,
		SUM( CASE WHEN i.i_category_id = 7 THEN 1 ELSE 0 END) AS clicks_in_7,
		SUM( CASE WHEN i.i_category_id = 8 THEN 1 ELSE 0 END) AS clicks_in_8,
		SUM( CASE WHEN i.i_category_id = 9 THEN 1 ELSE 0 END) AS clicks_in_9
	  FROM web_clickstreams_hdfs_parquet as w
	  INNER JOIN item as i ON (w.wcs_item_sk = i_item_sk
						 AND w.wcs_user_sk IS NOT NULL)
	  GROUP BY w.wcs_user_sk
	) AS q
	INNER JOIN customer as c ON q.wcs_user_sk = c.c_customer_sk
	INNER JOIN customer_demographics as cd ON c.c_current_cdemo_sk = cd.cd_demo_sk;
GO

-- Inspect top 100 rows
--
SELECT TOP(100) * FROM web_clickstreams_hdfs_book_clicks;
GO

-- Create the training stored procedure
CREATE OR ALTER PROCEDURE [dbo].[train_book_category_visitor_per_credit_rating_python]
(@model_name varchar(100))
AS
BEGIN
	DECLARE @model varbinary(max)
		, @model_native varbinary(max)
		, @input_query nvarchar(max)
		, @train_script nvarchar(max)
		
-- Set the input query for training. We will use 80% of the data.
	SET @input_query = N'
SELECT TOP(80) PERCENT SIGN(q.clicks_in_category) AS book_category
		, q.college_education
		, q.male
		, q.clicks_in_1
		, q.clicks_in_2
		, q.clicks_in_3
		, q.clicks_in_4
		, q.clicks_in_5
		, q.clicks_in_6
		, q.clicks_in_7
		, q.clicks_in_8
		, q.clicks_in_9
		, q.cd_credit_rating
	FROM web_clickstreams_hdfs_book_clicks as q
';
	-- Training R script that uses rxLogit function from RevoScaleR package (Microsoft R Server) to generate model to predict book_category click(s).
		SET @train_script = N'
# build classification model to predict book_category
import pickle
import pandas as pd
from revoscalepy import rx_logit, rx_serialize_model

# initialize the model variables returned to SQL Server
model = bytes()
model_native = bytes()

logit_model = rx_logit(formula = """
	book_category ~ college_education + male +
	clicks_in_1 + clicks_in_2 + clicks_in_3 + clicks_in_4 + clicks_in_5 +
	clicks_in_6 + clicks_in_7 + clicks_in_8 + clicks_in_9
""", data = indata
   , report_progress = 0);

model_native = rx_serialize_model(logit_model, realtime_scoring_only = True)
model = pickle.dumps(logit_model)
model_output = {"model_name": [model_name + " - " + indata["cd_credit_rating"].iloc[0]], "model": [model], "model_native": [model_native]}
OutputDataSet = pd.DataFrame(data=model_output, columns = ["model_name", "model", "model_native"])
';

	-- Generate sales model using Python script with the book clicks stats for each user
	-- Additionally we train a separate model for users in a specific credit rating
	-- The @input_data_partition_by_columns provides the execution of the Python script per partition (credit rating in this case)
	DELETE FROM sales_models WHERE model_name LIKE CONCAT(@model_name, '%');

	INSERT INTO sales_models (model_name, model, model_native)
	EXECUTE sp_execute_external_script
		  @language = N'Python'
		, @script = @train_script
		, @input_data_1 = @input_query
		, @input_data_1_name = N'indata'
		, @input_data_1_partition_by_columns = N'cd_credit_rating'
		, @params = N'@model_name varchar(100)'
		, @model_name = @model_name;
END;
GO


-- Step #1
-- Train the book category prediction model:
--
DECLARE @model_name varchar(100) = 'category_model_per_credit_rating (Python)';
EXECUTE [dbo].[train_book_category_visitor_per_credit_rating_python] @model_name;
SELECT * FROM sales_models WHERE model_name LIKE CONCAT(@model_name, '%');
GO

-- Step #2
-- Predict the book category clicks for new users based on their pattern of 
-- visiting various categories in the web site
-- We will use the model trained for each credit rating
--
DECLARE @model_name varchar(100) = 'category_model_per_credit_rating (Python)', @credit_rating varchar(50) = 'Low Risk';
DECLARE @sales_model varbinary(max) = (SELECT model_native FROM sales_models WHERE model_name = CONCAT(@model_name, ' - ', @credit_rating));
SELECT TOP(100)
      w.wcs_user_sk
	, p.book_category_Pred as book_click_prediction
	, w.college_education as [College Education]
	, w.clicks_in_1 AS [Home & Kitchen]
	, w.clicks_in_2 AS [Music]
	, w.clicks_in_3 AS [Books]
	, w.clicks_in_4 AS [Clothing & Accessories]
	, w.clicks_in_5 AS [Electronics]
	, w.clicks_in_6 AS [Tools & Home Improvement]
	, w.clicks_in_7 AS [Toys & Games]
	, w.clicks_in_8 AS [Movies & TV]
	, w.clicks_in_9 AS [Sports & Outdoors]
  FROM PREDICT(MODEL = @sales_model, DATA  = web_clickstreams_hdfs_book_clicks as w) WITH ("book_category_Pred" float) as p
 WHERE p.book_category_Pred <> SIGN(w.clicks_in_category)
   and w.cd_credit_rating = @credit_rating;
GO