We have released a new version of ML Studio, version 2.2. As usual, it contains multiple improvements. This release is focused on the improved handling of class-imbalanced data sets.
With version 2.2 we are introducing:
- Random Downsampling
- Model Calibration
- Cluster Creation Classifier
This feature is available for sample operations. The method reduces the class balance of a data set to the specified ratio by randomly removing records with a specific binary value while retaining records with the other binary value. You can read more about how it works here.
Use the Model Calibration feature to recalibrate a model's prediction probabilities to reflect a different class distribution than that of the training distribution. Model Calibration can be used in combination with Random Downsampling to reduce training time on highly class-imbalanced data sets while retaining strong calibration. Read more about Model Calibration here.
Cluster Creation from Classifier
Clustering models can now be generated directly from the Model Actions panel from any classification model. When generated via this feature, the Cluster Creation form is auto-populated with the model’s hyper-parameters. An applicable Cluster/Search specification data file is still required.
Several bugs were fixed, including:
- The Bias Detection Report now respects the maximum digits parameter from the Admin page.
- simSearch / simRecommend batch queries were not always listing the correct state after completion.
- Using the UI “Execute Query” button with all fields blank was resulting in an error in some model types.
This release also contains additional minor bug fixes and security updates.