Customizing PAI-Rec recommendation algorithms

Prerequisites

Before you begin, complete the following:

• Activate PAI. For details, see Activate PAI and create a default workspace.

• Create a Virtual Private Cloud (VPC) and a vSwitch. For details, see Create a VPC with an IPv4 CIDR block.

• Activate PAI-FeatureStore. For details, see the Prerequisites section of Create a data source. Hologres activation is not required. For the data source, select FeatureDB. For details, see Create an online data source: FeatureDB.

• Activate MaxCompute and create a MaxCompute project named project_mc. For details, see Activate MaxCompute and Create a MaxCompute project.

• Create an Object Storage Service (OSS) bucket. For details, see Create a bucket.

• Activate DataWorks:

  • Create a DataWorks workspace. For details, see Create a workspace.

  • Purchase a Serverless resource group for DataWorks. For details, see Use a Serverless resource group. This resource group synchronizes data for PAI-FeatureStore and runs eascmd commands to create and update the PAI-EAS service.

  • Configure DataWorks data sources:

    • Create and bind an OSS data source. For details, see Data Source Management.

    • Create and bind a MaxCompute data source. For details, see Bind a MaxCompute compute engine.

• Create a FeatureStore project and feature entities. Skip this step if you use a Serverless resource group. If you use a dedicated resource group for DataWorks, you must install the FeatureStore Python SDK on the resource group. For details, see Step 2: Create and register a FeatureStore and Install the FeatureStore Python SDK.

• Activate Realtime Compute for Apache Flink. For details, see Activate Realtime Compute for Apache Flink. Note: Set the storage type to OSS bucket, not Fully Managed Storage. Make sure that the OSS bucket for Flink is the same as the one configured in PAI-Rec. Flink records real-time user behavior data and computes real-time user features.

• If you later choose to use EasyRec (TensorFlow framework), training runs on MaxCompute by default.

• If you later choose to use TorchEasyRec (PyTorch framework), training runs on PAI-DLC by default. To download MaxCompute data on PAI-DLC, you must activate Data Transmission Service (see Purchase and use a dedicated resource group for Data Transmission Service).

1. Create a PAI-Rec instance

1. Log in to the PAI-Rec homepage and click Buy Now.

2. On the PAI-Rec instance purchase page, configure the following parameters and click Buy Now.

   Parameter	Description
   Region and availability zone	The region where your service is deployed.
   Service Type	For this solution, select Standard Edition and enable the Recommendation Solution Customization feature.

3. Log in to the PAI-Rec console. In the upper-left corner of the top menu bar, select a region.

4. In the navigation pane on the left, choose Instance list. Click the instance name to go to the instance details page.

5. In the Procedure section, click Cloud Service Configuration. This takes you to the System Configuration > Cloud Service Configuration page. Click Edit, configure the following parameters, and then click Exit.

   Resource configuration

   Parameter	Description
   Modeling
   Machine Learning Platform for AI Workspace	Enter your default PAI workspace.
   DataWorks Workspace	Enter the automatically generated DataWorks workspace.
   MaxCompute Workspace	Enter your MaxCompute project.
   OSS Bucket	Select your OSS bucket.
   Engine
   Real-time Recall Engine	For Whether to use PAI-FeatureStore, select Yes.
   Real-time Feature Query	For Whether to use PAI-FeatureStore, select Yes.

6. In the navigation pane on the left, choose System Configuration > Permissions. On the Services tab, verify that each cloud product is correctly authorized.

2. Clone the public dataset

1. Synchronize tables

This solution accepts input data in two ways:

1. Clone data for a fixed time window from the pai_online_project project. This method does not support routine scheduling.

2. Use a Python script to generate data. You can run a task in DataWorks to generate this data for a specific period.

To schedule daily data generation and model training, use the second method. Deploy the specified Python code to generate the data. For more information, see the Generate data using code tab.

CREATE TABLE IF NOT EXISTS rec_sln_demo_user_table_v1(
 user_id BIGINT COMMENT 'Unique user ID',
 gender STRING COMMENT 'Gender',
 age BIGINT COMMENT 'Age',
 city STRING COMMENT 'City',
 item_cnt BIGINT COMMENT 'Number of created items',
 follow_cnt BIGINT COMMENT 'Number of follows',
 follower_cnt BIGINT COMMENT 'Number of followers',
 register_time BIGINT COMMENT 'Registration time',
 tags STRING COMMENT 'User tags'
) PARTITIONED BY (ds STRING) STORED AS ALIORC;
INSERT OVERWRITE TABLE rec_sln_demo_user_table_v1 PARTITION(ds)
SELECT *
FROM pai_online_project.rec_sln_demo_user_table
WHERE ds >= "${bizdate_100}" and ds <= "${bizdate}";
CREATE TABLE IF NOT EXISTS rec_sln_demo_item_table_v1(
 item_id BIGINT COMMENT 'Item ID',
 duration DOUBLE COMMENT 'Video duration',
 title STRING COMMENT 'Title',
 category STRING COMMENT 'Primary tag',
 author BIGINT COMMENT 'Author',
 click_count BIGINT COMMENT 'Total clicks',
 praise_count BIGINT COMMENT 'Total likes',
 pub_time BIGINT COMMENT 'Publication time'
) PARTITIONED BY (ds STRING) STORED AS ALIORC;
INSERT OVERWRITE TABLE rec_sln_demo_item_table_v1 PARTITION(ds)
SELECT *
FROM pai_online_project.rec_sln_demo_item_table
WHERE ds >= "${bizdate_100}" and ds <= "${bizdate}";
CREATE TABLE IF NOT EXISTS rec_sln_demo_behavior_table_v1(
 request_id STRING COMMENT 'Instrumentation ID/Request ID',
 user_id STRING COMMENT 'Unique user ID',
 exp_id STRING COMMENT 'Experiment ID',
 page STRING COMMENT 'Page',
 net_type STRING COMMENT 'Network type',
 event_time BIGINT COMMENT 'Behavior time',
 item_id STRING COMMENT 'Item ID',
 event STRING COMMENT 'Behavior type',
 playtime DOUBLE COMMENT 'Playback/Read duration'
) PARTITIONED BY (ds STRING) STORED AS ALIORC;
INSERT OVERWRITE TABLE rec_sln_demo_behavior_table_v1 PARTITION(ds)
SELECT *
FROM pai_online_project.rec_sln_demo_behavior_table
WHERE ds >= "${bizdate_100}" and ds <= "${bizdate}";

2. Configure dependency nodes

To ensure smooth code generation and deployment, add three SQL code nodes to your DataWorks project. Set the scheduling dependencies of these nodes to the root node of the workspace. Finally, publish the nodes. The procedure is as follows:

1. Hover over Create and choose Create Node > General-purpose > Virtual Node. Create three virtual nodes with the following resource configuration, and click Confirm.

   Resource configuration

   Parameter	Description	Example
   Node type	Select the node type.	Virtual Node
   Path	The storage path for the node.	Workflow/Workflow/General
   Name	A unique name for the node.	rec_sln_demo_user_table_v1 rec_sln_demo_item_table_v1 rec_sln_demo_behavior_table_v1

2. For each virtual node, set its content to select 1;. Then, click Configure Scheduling on the right and configure the following settings:

   • In the Time property section, set Rerun property to Rerun When Succeeded Or Failed.

   • In the Scheduling dependencies > Upstream dependencies section, enter your DataWorks workspace name, select the node with the _root suffix, and click Add.

     Apply these settings to all three virtual nodes.

3. Click the icon next to each virtual node to submit it.

3. Register data

To configure feature engineering, recall, and sorting algorithms in the custom recommendation solution, you must first register the three tables that have been synchronized to your DataWorks project. The procedure is as follows:

1. Log in to the PAI-Rec console. In the top menu bar, select a region.

2. In the left-side navigation pane, choose Instance List. Click the instance name to open the instance details page.

3. In the left-side navigation pane, choose Recommendation Solution Customization > Data Registration. On the MaxCompute Table tab, click Create Data Table. Add one user table, one item table, and one behavior table as described in the following table, and then click Import.

   Parameter	Description	Example
   MaxCompute project	Select your MaxCompute project.	project_mc
   MaxCompute table	Select the data tables synchronized to your DataWorks workspace.	user table: rec_sln_demo_user_table_v1 item table: rec_sln_demo_item_table_v1 behavior table: rec_sln_demo_behavior_table_v1
   data table name	Enter a custom name.	user table item table behavior table

4. Create recommendation scenario

Before you can configure a recommendation task, you must first create a recommendation scenario. For an explanation of recommendation scenarios and traffic IDs, see Basic Concepts.

In the left navigation pane, click Recommendation Scenario. Click Create Scenario, configure the new scenario according to the following table, and click Determine.

5. Configure an algorithm solution

To fully configure a real-world scenario, use the following recall and fine-grained ranking configurations:

• global hot recall: Selects the top-k most popular items from log data.

• global hot fallback recall: Uses Redis as a fallback to prevent empty responses from the recommendation API.

• grouped hot recall: Recalls items by groups (e.g., city, gender) to provide more accurate recommendations of popular items.

• etrec u2i recall: Uses the etrec collaborative filtering algorithm.

• Swing u2i recall (optional): Uses the Swing algorithm.

• cold-start recall (optional): Handles cold-start scenarios using the DropoutNet algorithm.

• fine-grained ranking: Uses MultiTower for single-objective ranking and DBMTL for multi-objective ranking.

You should enable algorithms such as vector recall or PDN recall only after the initial recall covers most of your candidates. Vector recall requires a vector recall engine. FeatureDB does not support vector recall, so this example does not configure it.

This article shows a sample configuration and deployment. To simplify the process, this example uses only global hot recall and the u2i recall strategy from RECommender (eTREC, an implementation of collaborative filtering) for the recall configuration, and fine-grained ranking for the ranking configuration. The steps are as follows:

1. In the left-side navigation pane, choose Recommendation Solution Customization > Solution Configurations. Select the scenario that you created, click Create Recommendation Solution, provide the following resource configuration, and then click Save and Go to Algorithm Solution Configuration.

   Keep the default values for unspecified parameters. For more information, see Data table configuration.

   Resource configuration

   Parameter	Description
   Solution Name	Enter a custom name.
   Scenario Name	Select the recommendation scenario that you created.
   Offline Data Source	Select the MaxCompute project associated with the recommendation scenario.
   DataWorks Workspace	Select the DataWorks workspace associated with the recommendation scenario.
   Workflow Name	The name of the workflow that DataWorks creates when you deploy the solution script. You can enter a custom name, such as Flow.
   StorageAPI configuration	For regions in the Chinese mainland, such as Beijing and Shanghai, you can select "StorageAPI", which is a pay-as-you-go Data Transmission Service. For regions outside the Chinese mainland, such as China (Hong Kong), Singapore, and Frankfurt, you must first purchase and use a dedicated resource group for Data Transmission Service. If a pay-as-you-go option is not available, you must purchase a subscription for Data Transmission Service. Then, refresh the page and select the name of your Data Transmission Service subscription. Add a parameter to the PAI-DLC TorchEasyRec training task in DataWorks, for example: -odps_data_quota_name ot_xxxx_p#ot_yyyy.
   slim_mode	If the DataWorks edition you purchased has a size limit on the code packages imported by Migration Assistant, you can use this feature and manually upload the code packages that exceed the size limit. For this solution, select No.
   OSS Bucket	Select the OSS Bucket associated with the recommendation scenario.
   Project	Select the FeatureStore project that you created. For the online data source, select FeatureDB.
   User Entity	Select the user entity for the FeatureStore project, which is named 'user'.
   Item Entity	Select the item entity for the FeatureStore project, which is named 'item'.

2. In the Configure Table node, click Add next to the target data table. Configure the partitions, events, features, and timestamps for the behavior log table, user table, and item table as detailed in the following sections. Click Next.

   You can keep the default values for any parameters that are not described. For more information, see Data Table Configuration.

   Behavior log table

   Configure the behavior log table settings to match your data. In this example, the behavior log contains core information such as the request ID, user ID, behavior page, behavior timestamp, and behavior category. If your table has richer data dimensions, classify them by user and item to simplify feature engineering.

   Parameter	Description	Example
   Behavior Table Name	Select the registered behavior table.	rec_sln_demo_behavior_table_v1
   Time Partition	The partition field of the behavior table.	ds yyyymmdd
   Behavior information
   Request ID	The ID that identifies each recommendation request in the log. This is typically a UUID generated by the application. This field is optional.	request_id
   Behavior Event	The field in the log that records the behavior event.	event
   Behavior Event Enumeration Values	The enumerated values for the behavior event, such as impression, click, add-to-cart, or purchase.	expr, click, praise
   Behavior Value	The value that indicates the depth of a behavior, such as a transaction price or viewing duration.	playtime
   Behavior Timestamp	The event time, as a Unix timestamp in seconds.	event_time
   Timestamp Format	Specifies the format of the behavior timestamp.	unixtime
   Behavior Scenario	The field specifying the event scenario, such as a home page, search page, or product details page.	page
   Scenario Enumeration Values	The scenarios to use for data processing. This enables scenario-specific feature engineering.	home, detail
   User information
   User ID	The user ID field in the behavior table.	user_id
   User Categorical Features	User categorical features from the behavior table, such as network type, operating platform, or gender.	net_type
   Item information
   Item ID	The item ID field in the behavior table.	item_id

   User table

   Parameter	Description	Example
   User Table Name	Select the registered user table.	rec_sln_demo_user_table_v1
   Time Partition	The time partition field of the user table.	ds yyyymmdd
   User information
   User ID	The user ID field in the user table.	user_id
   Registration Timestamp	The time the user registered.	register_time
   Timestamp Format	Specifies the format of the registration timestamp.	unixtime
   Categorical Features	Categorical fields from the user table, such as gender, age group, or city.	gender, city
   Numerical Features	Numerical fields from the user table, such as the number of created items or points.	age, item_cnt, follow_cnt, follower_cnt
   Tag Feature	The field name for the tag feature.	tags

   Item table

   Parameter	Description	Example
   Item Table Name	Select the registered item table.	rec_sln_demo_item_table_v1
   Time Partition	The time partition field of the item table.	ds yyyymmdd
   Item information
   Item ID	The item ID field in the item table.	item_id
   Author ID	The author of the item.	author
   Listing Timestamp	The field for the item's listing timestamp.	pub_time
   Timestamp Format	Specifies the format of the listing timestamp.	unixtime
   Categorical Features	Categorical fields from the item table, such as category.	category
   Numerical Features	Numerical fields from the item table, such as price, total sales, or number of likes.	click_count, praise_count

3. In the Configure Feature node, configure the parameters as described in the following table, click Generate Feature, set a feature version, and then click Next.

   After you click Generate Feature, the system derives various statistical features for users and items. This solution uses the default derived features, but you can edit them based on your business requirements. For more information, see Feature Configuration.

   Resource configuration

   Parameter	Description	Example
   Statistical Period	Use this parameter for batch feature generation. To avoid generating too many features, set the statistics period to 3, 7, and 15 days. This setting calculates statistics for users and items over the specified periods. If you have very few user behavior events, try 21 days.	3,7,15
   Behavior	Select the configured behavior events. The recommended order is expr (impression), click, and praise.	expr, click, praise

4. In the Retrieval Configuration node, click Add next to the target category, configure the parameters, click Confirm, and then click Next.

   This section describes several recall configuration methods. To get started quickly, you only need to configure global hot recall and etrec u2i recall. The other methods are provided for reference.

   Resource configuration

   Global hot recall

   Global hot recall generates a ranking of popular items based on click event statistics, where top_n represents the number of items in the ranking. To modify the popularity scoring formula or access event, generate the relevant code and deploy it to DataWorks.

   The scoring formula is click_uv*click_uv/(expr+adj_factor)*exp(-item_publish_days/fresh_decay_denom), where:

   • click_uv: For the same click-through rate (CTR), a higher number of clicks indicates greater popularity.

   • click_uv/(expr+adj_factor): The smoothed CTR, where click_uv is the number of unique users who clicked and expr is the number of impressions. The adjustment factor adj_factor is added to prevent the denominator from being zero and to adjust the CTR when the number of impressions is low. When impressions are few, the CTR approaches 1. Adding adj_factor moves the CTR away from 1, making it closer to the true CTR.

   • exp(-item_publish_days/fresh_decay_denom): This term penalizes older items. item_publish_days is the number of days since the item was published.

   In the configuration editor, set parameters such as Recall Model Name (for example, global_hot), recall time window (for example, 15 days), recall count (for example, 500), Exposure Behavior Event, Click Behavior Event, the Hot Score Formula switch, recall engine (for example, FeatureStore), and Version.

   Etrec u2i recall

   etrec is an item-based collaborative filtering algorithm. For details, see Collaborative filtering etrec.

   In the Edit Configuration dialog box, set the Recall Model Name (for example, etrec) and recall engine (for example, Hologres). In the U2I Behavior Weight section, you can add events and their corresponding weights, such as a weight of 0 for expr, 1 for click, and 3 for praise.

   Parameter	Description
   Training days	The number of days of behavior logs used for training. The default is 30. You can adjust this value based on the log volume.
   Recall count	The number of user-to-item recommendations to generate offline.
   U2I trigger	Items the user has interacted with (for example, by clicking, favoriting, or purchasing). This typically excludes items the user only viewed (impressions).
   Behavior time window	The number of days of behavior data to collect. The default is 15, which includes data from the most recent 15 days.
   Behavior time attenuation coefficient	A value between 0 and 1. A larger value indicates that past behaviors decay more rapidly and therefore have less weight in constructing the trigger_item.
   Trigger selection count	The number of item IDs per user to use for the Cartesian product with the I2I data generated by etrec. A value between 10 and 50 is recommended. A large value can generate an excessive number of recall candidates.
   U2i behavior weight	For exposure events, either do not set a weight or set the weight to 0. We recommend not setting exposure events, thereby skipping user exposure data.
   I2I model settings	etrec parameter settings. For details, see Collaborative filtering etrec. We recommend not setting a high value for the number of related items. After enabling the I2I model, configure the following parameters: set similarity calculation strategy (sim_type) to wbcosine (options: asymcosine, jaccard); set related item selection count (top_n) to 500; set maximum behaviors per user (max_bhv) to 500; set minimum behaviors per user (min_bhv) to 2; set calculation strategy (operator) to add (options: mul, min, max); set similarity calculation weight coefficient (weight) to 1; and set decay coefficient (alpha) to 0.5. After completing the configuration, click Confirm.

   Grouped hot recall

   You can set up rankings based on attributes such as city and gender to provide initial personalized recall. In the following example, a combination of gender and the bucket ID of a numerical feature is used for grouping.

   On the Edit Configuration page, set recall type to grouped hot recall, Recall Model Name to group_hot, recall time window to 15, and recall count to 500. Set exposure behavior event to expr and click behavior event to click. Enable the Hot Score Formula Settings switch, then set hot score exponent adjustment factor to 100, hot score freshness decay to true, and hot score freshness decay factor to 180. In the User Group Trigger section, click Add to add features: gender (with no bucket boundaries) and follow_cnt (with bucket boundaries of 1,5,10,20). In the Behavior Group Trigger section, add the feature net_type (with no bucket boundaries). Set recall engine to Hologres and Version to 1, then click Confirm.

   Swing u2i recall

   Swing is a method for calculating item relevance that measures item similarity based on the User-Item-User pattern. In the Edit Configuration dialog box, configure the following parameters: set recall type to swing u2i recall, Recall Model Name to swing, training days to 30, and recall count to 500. Set recall engine to Hologres, disable U2I trigger, set behavior time window to 15, decay coefficient to 0.2, and trigger selection count to 10. In the U2I Behavior Weight section, click Add to add behavior weights: set the weight for expr to 0, click to 1, and praise to 3.

   The parameters for the Swing I2I model include: related item selection count (top_n, example: 500), maximum clicks per user (max_click_per_user, example: 600), maximum users per item (max_user_per_item, example: 700), maximum time span (max_time_span, example: 1), adjustment coefficients (alpha1=5, alpha2=1, beta=0.3), item weight calculation method (norm_method, option: COUNT), and Version (version). When you are finished, click Confirm.

   Vector recall

   Two vector recall methods are provided, DSSM and MIND. For more information, see the following:

   • Recall Target Name: Typically indicates whether an item was clicked. Set this to is_click.

   • Recall Target Selection: Set to max(if(event='click', 1, 0)).

     You can refer to the following code:

     select max(if(event='click',1,0)) is_click ,...
     from ${behavior_table}
     where between dt=${bizdate_start} and dt=${bizdate_end}
     group by req_id,user_id,item

     Where:

     • ${behavior_table}: The behavior table.

     • ${bizdate_start}: The start date of the behavior time window.

     • event: The event field in the ${behavior_table} table. Select a value based on your specific schema.

     • is_click: The target name.

     The formulas for dimension calculation are as follows:

     EMB_SQRT4_STEP8: (8 + Pow(count, 0.25)) / 8) * 8
     EMB_SQRT4_STEP4: (4 + Pow(count, 0.25)) / 4) * 4
     EMB_LN_STEP8:    (8 + Log(count + 1)) / 8) * 8
     EMB_LN_STEP4:    (4 + Log(count + 1)) / 4) * 4

     Here, count is the number of feature enumeration values. Use the Log function when a feature has a large number of unique values.

   In the Edit Configuration panel, set Recall Model Name to dssm and Model Type to dssm. Disable the negative sampling strategy. Enable recall target settings and set target type to CLASSIFICATION. Set training days to 30 and embed_dim strategy to EMB_SQRT4_STEP4. Set recall engine to Hologres. For incremental training, select true and set the incremental training days to 1. For both Asynchronous Training and online inference mode, select false. Set interest count to 1 and enable Sample Weight.

   Set sample weight name (name) to sample_weight and sample weight expression (selection) to ln(sum(playtime) + 1). Scene data filtering (scene_values) is used to enter scene data filtering conditions. Separate multiple values with commas. Set Version (version) to 1.

   Cold-start recall

   This is a dual-tower recall model similar to DSSM, divided into a user tower and an item tower. DropoutNet is a recall model suitable for head, long-tail, and even brand-new users and items.

   On the Create Configuration page, set the following parameters: set recall type to cold-start recall and Model Type (model_type) to dropoutnet. Enable Recall Target Settings (label) and configure the recall target name as is_click, the recall target selection as max(if(event='click', 1, 0)), and the target type as CLASSIFICATION. Set training days (train_days) to 30 and embed_dim strategy to EMB_SQRT4_STEP4. Set recall engine to Hologres. Set incremental training to true and set the incremental training days to 1. Set both Asynchronous Training and online inference mode to false. Set Version to 1 and click Confirm.

   Global hot fallback

   Global hot fallback recall is similar to global hot recall. Its main purpose is to provide a fallback with a sufficient candidate set if the primary global hot recall engine fails. Therefore, its output is stored in Redis. This output contains only a single row of data. On the Edit Configuration page, set recall type to global hot fallback recall. Set fallback model name (model_name) to global_hot_supplement, recall time window (day_interval) to 15, and fallback recall count (supplement_top_n) to 500. Set exposure behavior event (exposure_event) to expr and click behavior event (click_event) to click. Disable the Hot Score Formula Settings (hot_score) toggle. Set redis data source name (redis_datasource) to redis and Version (version) to 1.

   Collaborative metric i2i recall

   The collaborative metric learning i2i recall model uses session click data to calculate item-to-item similarity. Configure the model parameters: set the recall model name to CoMetricLearningI2I. After enabling Recall Target Settings, set the recall target name to is_click, the recall target selection to max(if(event='click', 1, 0)), and the target type to CLASSIFICATION. For the remaining settings, set training days to 30, set embed_dim strategy to EMB_SQRT4_STEP4, set recall engine to Hologres, set incremental training to false, set incremental training days to 1, set Asynchronous Training to false, and set version to 1. After completing the configuration, click Confirm.

5. In the Configure Sorting Method node, click Add next to Fine-grained Ranking, configure the parameters as described below, click Confirm, and then click Next.

   Resource configuration

   The platform provides multiple ranking models. For more information, see ranking model. This section explains how to configure the ranking parameters for the DBMTL multi-objective ranking model.

   On the ranking parameter configuration page, set ranking type to Fine-grained Ranking, enter dbmtl for fine-grained ranking model name (model_name), enter playtime for filter field (exclude_field), set training days (train_days) to 30, select dbmtl for model type selection (model_type), select false for incremental training (is_incremental), set incremental training days (incremental_train_days) to 1, select false for asynchronous training (is_async), and set version (version) to 1.

   Click Create next to Refined Ranking Target Settings (labels) to add the following two labels:

   • Target 1: Set target name to is_click, set target expression to max(if(event='click',1,0)), and select classification for target type.

   • Target 2 (Note: The 'l' in 'ln' is a lowercase L): Set target name to ln_playtime, set target expression to ln(sum(playtime)+1), set target dependency to is_click, select regression for target type, and then click OK.

6. In the Generate Script node, click Generate Deployment Script.

   Important

   After the deployment script is generated, the system provides an OSS address that contains all the deployment files. You can save this address to run the deployment script manually later.

7. Once the script is generated, click Determine in the dialog box to go to the Recommendation Solution Customization > Deployment History page.

   If script generation fails, check the run logs to identify and fix the error, then regenerate the script.

6. Deploy the recommendation

Once the script is generated, you can deploy it to DataWorks in one of two ways.

7. Freeze nodes

After the backfill completes, freeze the tasks in the Operation Center (the three nodes in step 2.2) to prevent them from running daily.

In the DataWorks Operation Center, navigate to Auto Triggered Task O&M > Recurring Job. Search for the node (for example, rec_sln_demo_user_table_v1), select the node (Workspace.node name), and choose Suspend (Freeze).