4.1 is the latest version of PMML, the Predictive Model Markup Language. Because it is the latest version, 4.1 is also the best. Why so? It incorporates new language elements for modeling techniques not supported in PMML 4.0. It also takes the PMML standard a step forward when it comes to post-processing of model outputs.
It is then with pride that Zementis is announcing support for PMML 4.1 throughout its scoring products, including ADAPA on Site and ADAPA on the Cloud (Amazon and IBM SmartCloud) as well as UPPI for in-database scoring and UPPI for Hadoop.
We have also updated the PMML Converter so that it now converts PMML files from older versions to version 4.1. The PMML Converter is actually part of ADAPA and UPPI and is responsible for converting, correcting, and validating PMML files when these are uploaded for scoring.
Our support for PMML 4.1 includes:
1) Scorecards (including support for reason or adverse codes and point allocation for complex attributes)
Predictive models contain a set of input attributes that are used to predict a certain target attribute. This prediction is seen as an assessment about a prospect, a customer, or a scenario for which you want to predict an outcome based on historical data. In a scorecard, the set of attributes selected during model development is readily available for inspection, as well as the scores associated with them. These partial scores are then summed so that an overall score is obtained for the target attribute.
If you would like to read more about the new PMML element for Scorecards, make sure to check out the paper we published at SIG KDD 2011 together with FICO. You can download it HERE.
With the enhanced post-processing capabilities offered by PMML 4.1, you can now represent business decisions, together with thresholds, directly in the language's Output element. This new 4.1 feature also allows for model outputs to be generically transformed with the use of any of the transformations and built-in functions made available in PMML 4.0 for pre-processing of input fields. ADAPA and UPPI also allow for pre-processing results to be used together with model outputs during post-processing.
3) Multiple Models
PMML 4.1 offers a more powerful and yet simpler way for the expression of multiple models. ADAPA and UPPI benefit from that. Our tools are now capable of scoring all the different multiple model types supported in PMML. These include model segmentation, composition, chaining and ensemble, which includes Random Forest Models.
4) Is the model scorable?
As part of PMML 4.1, a new attribute "isScorable" was added to all modeling elements. If this attribute is set to "false", a model is basically determined to be non-scorable and ADAPA and UPPI will ignore it. The isScorable flag was added to the standard as a way to flag models NOT destined to production deployment, but that are nonetheless part of the model building cycle. Non-scorable models are important to ensure best practices and transparency among peers.
5) New built-in functions
PMML 4.1 added three important built-in functions to standard. These are: median, product and lowercase. For a list of all PMML built-in functions, click HERE. Note that all these functions are supported by ADAPA and UPPI.
ADAPA and UPPI can score a myriad of predictive models and return back predictions in the form of scores, business decisions, probabilities, pseudo-probabilities, ids, confidence values, etc. With this latest release, our products can now also return derived fields computed prior to modeling as well as any output fields computed as a result of post-processing. In this way, ADAPA and UPPI can be used not only for model deployment and execution, but also for data analysis and processing before model training.
If you have any questions about PMML 4.1 and the features support in our products, please make sure to contact us.