Build predictions

The fourth data operation of this tutorial will consist in analyzing our data with a machine learning model.

🗺️ Overview

During this step, we will create a clustering machine learning model using BigQuery ML. Then, we will aggregate all our data into one BigQuery table and use our new model to analyze it.

🤖 Create your machine learning model

1. Go to the BigQuery web UI in the Cloud Console.

2. In the navigation panel, in the Resources section, select your project and your dataset.

3. Enter the following SQL query in the Query editor text area:

   ###################################################
   # Train Clustering Model in your Big Query Console
   ###################################################
   
   CREATE OR REPLACE MODEL my-gbq-dataset.model_clustering_store_iowa_liquor
   OPTIONS(model_type='kmeans', num_clusters=5, standardize_features = true) AS
   
   SELECT * except( store_number, store_name, volume_sold_liters, sale_q1, sale_q2, sale_q3, sale_q4)
   --FROM `my-gcp-project.my-gbq-dataset`
   FROM
   (
   
     with tmp as (
       SELECT store_number,
         store_name,
         EXTRACT(QUARTER FROM date) as quarter,
         case when(lower(category_name) like '%vodka%') then 'vodkas'
           when(lower(category_name) like '%whiskies%') then 'whiskies'
           when(lower(category_name) like '%rum%') then 'rum'
           when(lower(category_name) like '%liqueur%') then 'liqueur'
           when(lower(category_name) like '%tequila%') then 'tequila'
           when(lower(category_name) like '%schnapps%') then 'schnapps'
           when(lower(category_name) like '%gin%') then 'gin'
           when(lower(category_name) like '%cocktails%') then 'cocktails'
           when(lower(category_name) like '%brandies%') then 'brandies'
           when(lower(category_name) like '%spirit%') then 'spirit'
         else 'autre' end cat_alcool,
         sum(sale_dollars) as sale_dollars,
         sum(volume_sold_liters) as volume_sold_liters
       FROM dlk_demo_iowa_liquor_pda.sales_details
       where date >= '2019-01-01' and date < '2020-01-01'
       group by store_number,store_name,quarter,cat_alcool
     )
     select store_number,
       store_name,
       sum(sale_dollars) as sale_dollars,
       sum(volume_sold_liters) as volume_sold_liters,
   
       round(sum(case when(quarter = 1) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q1,
       round(sum(case when(quarter = 2) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q2,
       round(sum(case when(quarter = 3) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q3,
       round(sum(case when(quarter = 4) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q4,
   
       round(sum(case when(cat_alcool = 'vodkas') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_vodkas,
       round(sum(case when(cat_alcool = 'whiskies') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_whiskies,
       round(sum(case when(cat_alcool = 'rum') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_rum,
       round(sum(case when(cat_alcool = 'liqueur') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_liqueur,
       round(sum(case when(cat_alcool = 'tequila') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_tequila,
       round(sum(case when(cat_alcool = 'schnapps') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_schnapps,
       round(sum(case when(cat_alcool = 'gin') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_gin,
       round(sum(case when(cat_alcool = 'cocktails') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_cocktails,
       round(sum(case when(cat_alcool = 'brandies') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_brandies,
       round(sum(case when(cat_alcool = 'spirit') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_spirit
     from tmp
     group by store_number,store_name
   )

4. Click Run.

The query takes several minutes to complete. After the first iteration is complete, your model (sample_model) appears in the navigation panel of the BigQuery web UI.

You can observe the model as it's being trained by viewing the Training tab in the BigQuery web UI.

📄 Create your configuration files

Create the JSON file that configures the data pipeline operation

1. Access your tailer-demo folder.

2. Inside, create a folder named 4-Build-predictions for this new step.

3. In this folder, create a JSON file named tailer-demo-build-predictions.json for your data operation.

4. Copy the following contents into your file:

   {
     "configuration_type": "table-to-table",
     "configuration_id": "000099-tailer-demo-build-predictions",
     "short_description": "Build Tailer demo predictions",
     "account": "000099",
     "environment": "DEV",
     "activated": true,
     "archived": false,
     "start_date": "2019, 1, 23",
     "catchup": false,
     "schedule_interval": "None",
     "max_active_runs": 1,
     "task_concurrency": 3,
     "default_gcp_project_id": "my-gcp-project",
     "default_bq_dataset": "my-gbq-dataset",
     "default_write_disposition": "WRITE_TRUNCATE",
     "task_dependencies": ["iowa_liquor_agg_store >> store_clustering"],
     "workflow": [
       {
         "task_type": "sql",
         "id": "iowa_liquor_agg_store",
         "short_description": "Aggregate data for clustering ",
         "table_name": "iowa_liquor_agg_store",
         "sql_file": "iowa_liquor_agg_store.sql"
       },
       {
         "task_type": "sql",
         "id": "store_clustering",
         "short_description": "Affect every store to the right cluster",
         "table_name": "store_clustering",
         "sql_file": "store_clustering.sql"
       }
     ]
   }

5. Edit the following values: ◾ Replace my-gcp-project-id with the ID of the GCP project containing your BigQuery dataset. ◾ Replace my-gbq-dataset with the name of your working dataset.

Create the JSON file that triggers the workflow

1. Inside the 4-Build-predictions folder, create a file named workflow.json.

2. Copy the following contents into your file:

   {
     "configuration_type": "workflow",
     "configuration_id": "000099-tailer-demo-build-predictions-workflow",
     "environment": "DEV",
     "short_description": "Launch the Tailer demo model execution with BQ",
     "account": "000099",
     "activated": true,
     "archived": false,
     "authorized_job_ids": ["gbq-to-gbq|tailer-demo-build-predictions_DEV"],
     "target_dag": "tailer-demo-build-predictions_DEV",
     "extra_parameters": {}
   }

Create SQL files

1. Inside the 4-Build-predictions folder, create the following files:\

   ◾ iowa_liquor_agg_store.sql

   ◾ store_clustering.sql

2. Copy the following contents into the iowa_liquor_agg_store.sql file:

   with tmp as (
   
   SELECT store_number,
     store_name,
     EXTRACT(QUARTER FROM date) as quarter,
     case when(lower(category_name) like '%vodka%') then 'vodkas'
       when(lower(category_name) like '%whiskies%') then 'whiskies'
       when(lower(category_name) like '%rum%') then 'rum'
       when(lower(category_name) like '%liqueur%') then 'liqueur'
       when(lower(category_name) like '%tequila%') then 'tequila'
       when(lower(category_name) like '%schnapps%') then 'schnapps'
       when(lower(category_name) like '%gin%') then 'gin'
       when(lower(category_name) like '%cocktails%') then 'cocktails'
       when(lower(category_name) like '%brandies%') then 'brandies'
       when(lower(category_name) like '%spirit%') then 'spirit'
     else 'autre' end cat_alcool,
     sum(sale_dollars) as sale_dollars,
     sum(volume_sold_liters) as volume_sold_liters
   FROM dlk_demo_iowa_liquor_pda.sales_details
   where date >= '2019-01-01' and date < '2020-01-01'
   group by store_number,store_name,quarter,cat_alcool
   
   )
   
   select store_number,
     store_name,
     sum(sale_dollars) as sale_dollars,
     sum(volume_sold_liters) as volume_sold_liters,
   
     round(sum(case when(quarter = 1) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q1,
     round(sum(case when(quarter = 2) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q2,
     round(sum(case when(quarter = 3) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q3,
     round(sum(case when(quarter = 4) then sale_dollars else 0 end) / sum(sale_dollars) , 2) as sale_q4,
   
     round(sum(case when(cat_alcool = 'vodkas') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_vodkas,
     round(sum(case when(cat_alcool = 'whiskies') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_whiskies,
     round(sum(case when(cat_alcool = 'rum') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_rum,
     round(sum(case when(cat_alcool = 'liqueur') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_liqueur,
     round(sum(case when(cat_alcool = 'tequila') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_tequila,
     round(sum(case when(cat_alcool = 'schnapps') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_schnapps,
     round(sum(case when(cat_alcool = 'gin') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_gin,
     round(sum(case when(cat_alcool = 'cocktails') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_cocktails,
     round(sum(case when(cat_alcool = 'brandies') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_brandies,
     round(sum(case when(cat_alcool = 'spirit') then sale_dollars else 0 end) / sum(sale_dollars) , 2) as p_spirit
   
   from tmp
   group by store_number,store_name

3. Replace my-gbq-dataset with the name of your working dataset.

4. Copy the following contents into the store_clustering.sql file:

   ######################################################
   # Affect every store to the right cluster
   ######################################################
   
   SELECT * except(nearest_centroids_distance) 
   FROM ML.PREDICT(MODEL my-gbq-dataset.model_clustering_store_iowa_liquor, 
   (
   SELECT * except(sale_q1, sale_q2, sale_q3, sale_q4)
   FROM `my-gbq-dataset.iowa_liquor_agg_store`
   
   ))

5. Replace my-gbq-dataset with the name of your working dataset.

▶️ Deploy the data operation

Once your files are ready, you can deploy the data operation:

1. Access your working folder by running the following command:

   cd "[path to your tailer folder]\tailer-demo\4-Build-predictions"

2. To deploy the data operation, run the following command:

   tailer deploy configuration 000099-tailer-demo-build-predictions.json

3. To trigger the workflow, run the following command:

   tailer deploy configuration workflow.json

Your data operation is now deployed, which means all your data will shortly be aggregated and analyzed by the machine learning model you have created. The Evaluation tab of your model allows you to view data clustering and get some first insights out of your data.

Your data operation status is now visible in Tailer Studio.

✅ Check the data operation status in Tailer Studio

1. Access Tailer Studio.‌
2. In the left navigation menu, select Table-to-table.

3. In the Configurations tab, search for your data operation, 000099-tailer-demo-build-predictions. You can see its status is Activated.

4. Click the data operation ID to display its parameters and full JSON file, or to leave comments about it.