Docs

ContextSearch

Personalization Search

Model development

In personalization search scenarios, after you import a dataset, you can perform feature engineering and sample engineering on it to develop models. This topic describes how to develop models.

Prerequisites

A personalization search scenario has been created. For more information, see Create a scenario.

Data has been imported. For more information, see Import data.

Create a development task

Log on to the ContextSearch console.

In the top navigation bar, select the target project and region.

Select My Scene in the left navigation panel, then click the target personalized search push scenario.

Click the Icon on the top-right corner on the target personalized search and recommendation scenario page, then select the Model dDevelopment tab.

In the Development task section, click Create.

In the new model development task dialog, enter the task name and click Create.

tip

Task name rules are as follows:

Cannot start with a number or a hyphen (-).

Length must be between 1 and 128 characters.

Contain letters, digits, underscores (_), and hyphens (-).

Develop model

In the created development task, click Dataset, then select the added dataset from the dropdown list.

tip

You can preview data in the dataset in the Item data preview area.

To add a new dataset, click Add dataset. For detailed steps, see Add dataset.

In the created development task, click Feature processing, check the fields to use in the Whether to adopt column of each table, and configure the Feature type.

tip

The feature types of fields are as follows:

Sparse features: non-continuous features such as String or non-numeric fields, names, or properties.

Dense features: continuously varying features that must be numeric fields, such as price or visit count.

Variable-length sparse features: lists of sparse features, for example, the list of interacted items item_id_list.

Variable-length dense features: lists of dense features, such as the list of prices of interacted items item_price_list.

Note that fields like id are sparse features because they do not vary continuously; for example, id=3 does not represent a specific number, id=3 is not greater than id=2. Only numeric values that can be compared, like prices, can serve as dense features.

In the created development task, click Sample project, then configure the following parameters.

Parameter	Description
Positive sample	The expected values that meet the demand output can be represented by the same field indicating whether the expectation is met, for example, clicked=0/1 indicates whether the user clicked the item, and purchase=0/1 indicates whether the item was purchased.
Positive sample ratio	Enter the proportion of positive samples, unit: %
Fields	Select the field used to split the dataset from the dropdown list.
ratio	Set the ratio of training, validation, and test sets. The total ratio equals 10.

In the created development task, click Model training, then configure the following parameters.

Parameter	Description
Model type	Currently fixed to recall + fine ranking.
Recall model algorithm	DSSM is a fixed option. DSSM (Deep Structured Semantic Model) is a deep learning model used to learn the semantic matching relationship between two types of objects, such as users and items, or queries and documents. You can click Hyperparameter configuration to configure the algorithm’s hyperparameters. For details, see DSSM hyperparameters. You can also click Add and then select other models from the dropdown list.
ranking model algorithm	Currently supports only DeepFM. DeepFM (Deep Factorization Machine) is a recommendation model that learns both low-order feature combinations and high-order feature interactions. You can click Hyperparameter configuration to configure algorithm hyperparameters. For details, see DeepFM hyperparameters.

DSSM hyperparameters

Parameter	Type	Required	Description
batch_size	INTEGER	Yes	Data for each training batch.
dropout	FLOAT	Yes	Dropout rate to prevent overfitting (default value means randomly dropping 10% of neurons).
early_stop_patience	INTEGER	Yes	Early stopping patience: The default value means training stops if there is no improvement in validation metrics for 5 consecutive epochs.
embedding_dim	INTEGER	Yes	Embedding vector dimension for categorical features.
epochs	INTEGER	Yes	Training epochs.
fe_block_layer_dim	INTEGER	Yes	Hidden layer dimension of the feature extraction module (FE Block).
full_reg	BOOLEAN	Yes	Enable full regularization.
item_head	INTEGER	Yes	Number of output heads for the item tower.
learning_rate	FLOAT	Yes	learning rate.
seed	INTEGER	Yes	Random seed to ensure experiment reproducibility.
sparse	BOOLEAN	Yes	Enable sparse training?
tau	FLOAT	Yes	Temperature coefficient used for similarity scaling in contrastive learning / two-tower matching.
user_head	INTEGER	Yes	The number of output heads in the user tower.

DeepFM hyperparameters

Parameter	Type	Required	Description
adagrad	BOOLEAN	Yes	Enable AdaGrad optimizer.
batch_size	INTEGER	Yes	Data for each training batch.
dcn_low_rank_dim	INTEGER	Yes	DCN low-rank decomposition dimension, used to reduce the number of parameters in the crossing layer.
dcn_num_layers	INTEGER	Yes	DCN (Deep & Cross Network) cross layer quantity: 3 layers.
embedding_dim	INTEGER	Yes	Embedding vector dimension for categorical features.
epochs	INTEGER	Yes	Number of training epochs.
learning_rate	FLOAT	Yes	learning rate.
lr_decay_start	INTEGER	Yes	Step to start learning rate decay (0 means no decay).
lr_decay_steps	INTEGER	Yes	Total number of steps for learning rate decay (0 means no decay).
lr_warmup_steps	INTEGER	Yes	Learning rate warm-up steps (0 means no warm-up).
pin_memory	BOOLEAN	Yes	Enable pinned memory (accelerates GPU data transfer, specific to PyTorch).

Run task

Click Run at the top right corner of the development task.

In the run task dialog, configure the following parameters, then click Submit.

Parameter	Description
Resource type	Currently fixed to GPU.
Specification	Select the resource specification from the dropdown list.
Quantity	Select resource quantity.
Dispatch	Select the running mode. Valid values: Run once: run immediately once. Hourly: run once at the specified minute every hour within the specified date range. Daily: run once at the specified time every day within the specified date range.

tip

Resources used for running tasks are billed based on actual usage. Please refer to the bill for detailed charges.

Subsequent operations

View logs

After the task runs, you can view the running logs to understand the model’s running state and locate issues.

Click Running history at the top right corner of the page.

In the Task running history panel, click Logs to the right of the desired running record.

On the log query page, select filter criteria such as log level, start and end time, file type, and keywords; the system will automatically filter and display matching logs in the list below.

tip

The system saves logs for a maximum of 7 days.

Model evaluation

After the task starts running, you can click Run history to view the task status. For example, you can view model evaluation after completing model training.

In the Run history panel, click Configuration information next to the target run record.

In the pop-up window, select the Model evaluation tab.

On the Model evaluation tab, view the recall training model metrics and re-ranking model training metrics for the running tasks.

Last updated: 2026.09.09 17:26:51