I got a chance to catch up with Zementis. I last got an update from them on the product back in May 2010 when they released 3.0 with their support of Drools as well as PMML. Zementis is reporting good momentum with large customers such as the US Army and Verizon as well as a number of strategic partnerships – Revolution Analytics (deployment of R models, reviewed here), KNIME (another open source data mining tool, reviewed here), Datameer (allowing analytics execution on Hadoop), Predixion (cloud based analytic modeling, reviewed here) and EMC Greenplum (analytic appliance) as well as Dymatrix (European partner for consulting and implementation for their customers there).
Since ADAPA 3 (reviewed here), ADAPA has added support for some new algorithms such as association rules, cox regression models, rule set models and broader support for PMML 4.0. More support for Drools is coming, with ruleflow and other non-DRL artifacts to be supported in a future release. ADAPA is available on the amazon.com compute cloud as well as on-premise. Customers like the flexibility of having both deployment options – some move from one to the other, some use both.
Their new product is the Universal PMML Plugin for in-database analytics, released with EMC Greenplum as a first database partner. This takes the core PMML execution engine from Zementis and moves it into EMC Greenplum so that models can be imported from PMML to create UDFs in the database. The executing models leverage the massively parallel infrastructure in EMC Greenplum to generate phenomenal performance (to be published soon) on the order of 1M scores/second. This joint development effort is being sold by both companies and is apparently generating a great deal of interest.
Zementis’ leadership in the PMML community continues too. They have launched their PMML conversion tool to the community to convert and validate everyone’s PMML. This is helping make PMML a more viable standard as it gives people a tool to manage the variations in the standard. Their execution engine, ADAPA, uses the same convertor. PMML is continuing to evolve too. PMML has not traditionally supported the pre-processing required by a model – variable transformations – but the newest version does. Based on this Zementis released PMML Transformations Generator – an interactive user interface to construct all the pre-processing elements for a model and add them to a PMML file. This is available free on their website and allows those using PMML to export models to quickly add their transformation logic to the PMML file. In addition some vendors are adding transformations to their PMML output and the Zementis folks see this growing rapidly with folks like KNIME adding full support for pre-processing in their PMML export. I am looking forward to increased support here as it makes it easier and quicker to deploy models to rules-based systems that support PMML. And of course they released a great little book on PMML to which I wrote an introduction.
PMML meanwhile is growing. More vendors have joined the Data Mining Group, many more people are participating in online communities about PMML. The growth in predictive analytics and the move towards a more industrial mindset is definitely driving interest in PMML. PMML also allows modelers to select modeling tools without worrying about the deployment piece – if your deployment environment supports PMML then you can use any tool that generates PMML.
Good to see PMML and its supporters continuing to grow. I look forward to seeing more database PMML plugins.
Don’t forget the Decision Management Technology Map