Start model training

Prerequisites

You have configured private set intersection and successfully run a virtual wide table.

Step 1: Create a model project

Before you start model training, you must create a model project.

1. Log on to the Multi-Party Security Modeling console.

2. In the navigation pane on the left, click Model Development > Model Training.

3. Click Create Model Project and configure the following parameters.

   1. Select a development mode. Only DAG Visualization is supported.

   2. Select a template.

      You can create a blank template or select an existing template and modify the modeling flow.

   3. Configure basic information.

      Parameter	Description
      Project Name	The name of the model project. Only letters, digits, and underscores (_) are supported.
      Description	The description of the model project. The system fills this in automatically.

4. Click Done. The model project is created and the visualization canvas opens automatically.

Step 2: Configure components

After you create a model project, you can configure its components.

1. Log on to the Multi-Party Security Modeling console.

2. In the navigation pane on the left, click Model Development > Model Training.

3. Click Edit to the right of the model project to open the visualization canvas.

   The visualization canvas consists of the following sections:

   • Flow

     The flow displays the model training process for the current project. You can select an existing model training flow, or manually add or delete flows.

   • Components

     The component list on the left side of the page contains algorithms. You can select components from this list and drag them onto the canvas.

   • Canvas

     The modeling canvas is in the center of the page. You can add components to the canvas and connect them to define their input and output relationships. This lets you customize the model training flow.

   • Parameter settings

     When you click a component on the canvas, the Parameter settings panel for that component appears on the right. In this panel, you can adjust the model training parameters to optimize the model.

   • Task list

     This list displays all model training tasks in the current model training flow. You can view the details of any task in the list.

Step 3: Create a training task

After you customize a complete model training flow on the canvas, you can create a model training task.

1. Right-click a component on the canvas, select an execution mode, and then execute the task.

   A model training task can be executed in one of the following three modes:

   • Execute from this node

   • Execute this node (single-step execution)

   • Execute to this node

2. View the task execution status.

   In the Task list, you can view the execution status of all tasks.

   Note

   Double-click a component on the canvas to open the task execution details page. This page displays information about the most recent task that includes the component.

3. View the training status of the components.

   Click Details to the right of a task to view the training status of each component in the task.

   Note

   In Multi-Party Computation (MPC) modeling, you cannot view the plaintext data generated after a component is executed. You can view the statistical values and evaluation results generated by exploration or evaluation components, such as correlation coefficients or model evaluation metrics.

4. View the training results of a component.

   On the task details page, click a component, and then click Output Table below to view the training results of the component.

5. Save the task.

   When you run multiple tasks on the same canvas, the output table of a new task overwrites the output table of the previous task. To save the output table of a specific task, go to the Task list, and click Click to Save next to the task.

   Note

   • You can save only the most recent task, and only if it has been executed successfully.

   • If a task has been overwritten, you can reproduce its components and parameters on the canvas, re-execute the task, and then save it.

Step 4: Submit the model

After you debug and finalize the model, you can submit it. Before you submit the model, you must save the task that generated it.

1. In the Task list below the visualization canvas, click Submit Model to the right of the task.

2. On the task execution details page, click Submit Model in the upper-right corner.

3. On the Select Components to Submit and Publish page, select the components that contain the model file and the pre-processing components from the current task, and then click Next.

   • Model file

     The model file is saved in the model training component. When you submit and publish components, you must select a training-type component, such as a logistic regression or XGBoost regression component.

   • Pre-processing components

     The Multi-Party Security Modeling console lets you package and submit pre-processing components with the model for deployment.

4. On the Submit Model page, enter the model information and click Submit.

   For models in the same scenario, use the same service name. Use the version number and service description to differentiate between models.