This topic describes how to use models trained by Machine Learning Designer to make predictions in the production environment on new data.
After you deploy a model to Elastic Algorithm Service (EAS) as a service, you can use the model to make predictions on new data. Machine Learning Designer supports two types of prediction services: real-time predictions and batch predictions. You can choose between these types based on your requirements on prediction timeliness. Machine Learning Designer supports both types of prediction services.
Real-time predictions
Deploy a single model as an online service
You can deploy models as online services to implement real-time prediction. After a pipeline is successfully run, you can deploy a model generated by the pipeline as an online service to EAS with a few clicks. Push-button deployment is available for trained Predictive Model Markup Language (PMML), AlinkModel, and XGBoost models. You can deploy a single model each time. You can also manually export trained PMML model files to your computer, upload the files, and then deploy the models as online services to EAS. Models in the Parameter Server (PS) format require some preparation before you can deploy them as online services in this way.
Deploy a pipeline as an online service
You can deploy specific pipelines as online services to implement real-time prediction. Specifically, you can use Alink components to create a batch data-processing pipeline that combines data preprocessing, feature engineering, and model prediction, and then deploy the pipeline as an online service. This can be achieved with a few clicks after packaging the pipeline as a batch model.
Periodically update online model services
After you deploy a model trained by Machine Learning Designer as an online service, you can connect the Update EAS Service component as a downstream node of the component that generates the model to update the online service. You can also submit a pipeline to DataWorks to schedule it as a periodic task. This way, the model service that is deployed from the pipeline is automatically updated at specific points in time.
Batch predictions
You can add prediction components to a pipeline to implement batch prediction by using Machine Learning Designer. Then, submit the pipeline to DataWorks and schedule it as a periodic task to implement automated prediction at specific points in time. For more information, see Implement batch prediction.