This page contains release notes for each version of the Prediction API.
Current version: 1.6
Deprecated versions (these versions will cease to work on 1st October 2013 - please upgrade your models before then):
Released June 2013
- Change in ownership of models: Models are now resources belonging to your Google Developers Console project - this allows sharing with other users in your team, or service accounts used by your application. In previous versions models could only be used by the user that created them.
- Training data size limit raised to 2.5 GB.
Important: with the release of V1.6 of Prediction API, all previous versions are deprecated and will no longer be supported after 1 October 2013. Please follow these steps to update your models to be compatible with v1.6. If no action is taken, existing models will become inaccessible and deleted after 1 October 2013.
Released October 2012
- Support for updating regression models.
- Support for training directly from instances passed in prediction.trainedmodels.insert (instead of a file in Google Storage).
Released March 2012
- Two new methods added: List and Analyze.
dataAnalysisare now returned with prediction.trainedmodels.analyze.
trainingCompleteis now returned with prediction.trainedmodels.get.
Released October 2011
- Support for PMML data preprocessing
- ; the REST URLs have changed for V1.4, as well as the request data syntax for most methods and the resource descriptions.
- Better training status notifications
- Training status (
trainedmodels.get) now returns HTTP 404 NOT FOUND until training completes successfully the first time, rather than
Released August 2011
- Improved error messages
- New training resource properties:
- modelInfo.confusionMatrix - A confusion matrix for analyzing model accuracy in categorical models.
- training.utility - A way to add an arbitrary cost associated with mislabeling an item of a specific category.
Released May 2011
A non-breaking version update that added the following features:
Released April 2011
- New JSON format for all requests and responses; breaking changes on all GET and POST JSON objects.
- Both Windows and UNIX newline characters supported in training data.
- Mean squared error values now returned for regression models.
- New terms of service announced.
- Training status request returns error messages for failed training requests.
- The "text", "numeric", and "mixture" values for prediction queries are deprecated. Only type now supported is "csvInstance".
- ClientLogin is now deprecated in favor of OAuth.
- The key parameter is now required in all requests.
- Categorical prediction scores normalized to fall between 0.0 and 1.0 Note: You must retrain your model in order for scores to be scaled to fall between 0-1.
Released September 2010
- Support for mixed text and numeric data
- Support for regression data
- Category queries return all possible categories, with scores for each.
- Ability to delete training models