Exporting model training data

You can export the model analysis data in the Compare and Analyze tabs in an experiment. Exporting the data exports it to the Qlik Cloud platform in the dedicated space, where you can further analyze it in Qlik Sense apps.

After the data is exported, you can import it into Qlik Sense apps in the following ways:

• Load the datasets into apps using the Data manager and Data catalog interfaces.

• Use Data load editor in the app to load the data using scripting.

• Create scripts and data flows to further transform and store the data to new files, which can be loaded into Qlik Sense apps.

Available formats

The model training data can be exported in the following formats:

• Parquet (default)

• CSV

• QVD

Exporting analysis data for the entire experiment

The following are available:

• Model metrics: Exports performance metrics for all models trained in the experiment. The performance metrics are generated by testing the trained models against the automatic holdout data. The dataset also includes performance metrics generated by testing the trained models against the training data itself.

• Hyperparameters: Exports data for the hyperparameters that were used when training the model.

Do the following:

1. Open the Compare tab in an ML experiment.

2. Click Export data to catalog above the embedded analysis.

3. Use the check boxes to select or clear options as needed.

4. As needed, edit the default dataset paths, including folders and the file name. Separate folders with / characters.

   For more information about folder references, see Tips for folder references in paths.

   

5. Select the output format for the data.

6. Select a space where the exported data will be stored.

7. Click the button to export the datasets.

Exporting analysis data for an individual model

The following are available:

• Prediction data: Exports the prediction data for the predictions the model has created on the automatic holdout data. For classification models, probabilities for each class are included.

• SHAP and test data: Exports the SHAP data calculated by the model on the automatic holdout data. The actual feature values for the automatic holdout data are also included in the dataset.

• Feature metadata: Exports a dataset with the date type and feature type for each feature used to train the model.

Do the following:

1. In the Analyze tab in an ML experiment, select a specific model, or click Analyze next to a model from another view.

2. Click Export data to catalog above the embedded analysis.

3. Use the check boxes to select or clear options as needed.

4. As needed, edit the default dataset paths, including folders and the file name. Separate folders with / characters.

   For more information about folder references, see Tips for folder references in paths.

   

5. Select the output format for the data.

6. Select a space where the exported data will be stored.

7. Click the button to export the datasets.

Examples:  

• A model metrics dataset with the following name and location: Model Performance/MyExperiment. This could store a dataset MyExperiment to a folder ModelPerformance within a space.

• A feature metadata dataset with the following name and location: Model Performance/Candidate Models/v01_LOGC_00_00. This could store a dataset v01_LOGC_00_00 to a folder structure Model Performance/Candidate Models in a space.