TL;DR
This report captures a 2009 "state of the union" for data mining, where industry titans reached a consensus that open standards—specifically PMML—and the then-nascent cloud computing were the keys to democratizing predictive analytics. It marks the transition from manual, vendor-locked model deployment to a future of automated, scalable, and interoperable "scoring" in the cloud.
What Problem Does This Paper Address?
In 2009, the data mining industry faced a massive "deployment gap." While researchers were developing sophisticated algorithms, moving those models into production was a significant challenge. The panel identifies that roughly 75–80% of the effort in analytics was spent on model preparation and dealing with "silos."
The primary challenges addressed were:
- Vendor Lock-in: Models built in one tool (like SAS or SPSS) could not easily be run in another environment or embedded into operational databases.
- Scalability Limits: Traditional computational models were "memory-resident," meaning they couldn't handle the "explosion in data volumes and dimensions" that was beginning to occur.
- High Barriers to Entry: Advanced analytics required massive up-front investments in hardware and specialized software licenses, pricing out small and medium businesses (SMBs).
Methods
As a panel report, this document does not present a new algorithm. Instead, it synthesizes the strategic approaches of nine industry leaders regarding two primary "methods" for modernization:
- PMML (Predictive Model Markup Language): The panel champions PMML (then at version 4.0) as the "SQL of data mining." It is an XML-based standard that allows a model to be defined in a tool-agnostic way, enabling a model trained in one application to be "scored" (executed) in another.
- Cloud Computing & SaaS: The panel explores the shift from perpetual on-premise licensing to Software-as-a-Service. They discuss the technical feasibility of moving "scoring" to the cloud versus the more difficult task of "training" models in the cloud due to data transfer bottlenecks.
Key Insights and Ideas
- The Democratization of Analytics: The panel predicted that cloud computing would lower the total cost of ownership so that the "community at large can focus on the essence of algorithms" rather than proprietary plumbing.
- Scoring vs. Training: A recurring insight is that the cloud is immediately ready for scoring (applying a model to new data), but training (building the model) remains a challenge due to the "gravity" of large datasets and the bandwidth costs of moving them to the cloud.
- The Parallelism Crisis: Mike Hoskins (Pervasive Software) noted a "frightening specter": while multi-core hardware was becoming common, the data mining software industry was largely unprepared to exploit it, still relying on single-threaded code.
- Privacy-Preserving Mining: Christoph Lingenfelder (IBM) highlighted that distributed mining (the precursor to federated learning) would allow organizations to find commonalities in sensitive data (like cancer cases) without ever actually sharing or merging the raw data.
- The End of Perpetual Licenses: The panel correctly predicted a shift toward "variable" licensing—subscriptions, transaction-based pay, or data-volume-based pricing—aligning with the elastic nature of the cloud.
Key Results
The report provides qualitative consensus rather than quantitative metrics:
- De-facto Standardization: PMML was officially recognized by all major vendors (IBM, SAS, SPSS, etc.) as the necessary standard for model exchange.
- Cloud as the Disruptor: Cloud computing was identified as the single most disruptive technology for the near term, specifically for its ability to provide "supercomputer-class processor grids" at a low cost.
- Interoperability Benefits: The panel concluded that open standards would reduce the risk for integrators and allow "best-of-breed" stacks where users could pick the best modeling tool and the best deployment engine independently.
Limitations and Future Work
- The "Semi-Open" Problem: Erik Marcadé (KXEN) warned that PMML lacked a formal compliance process. Vendors could claim to be "PMML compliant" while still having subtle incompatibilities.
- Data Preprocessing Gap: Rich Holada (SPSS) noted that while PMML 4.0 was great for the model, it didn't yet fully capture the source data transformations or the "actual training process," leaving a gap in full model lineage.
- Explorative Analytics: Michael Berthold (KNIME) argued that tools still required users to "know the question before they start." He suggested future work should focus on more flexible, visual workflows that allow for "explorative" discovery.
- Security and Bandwidth: The panel acknowledged that "disks of data" might still need to be physically shipped to cloud providers in cases where internet bandwidth couldn't handle the scale of 2009-era data volumes.
Summary updated