Coming Soon to a Mainframe Near You: Machine Learning, Part 1
So far the mainframe big data story has been very useful, but pretty tame: logs for operational intelligence, improved cybersecurity, improved retention period, fancier dashboards. Here’s betting that it’s going to get much more interesting — and probably already is in some shops.
We’re talking about artificial intelligence (AI) — and specifically machine learning (ML). ML is a discipline that Google has fully embraced. As Steven Levy wrote, Google is “remaking itself as a ‘machine learning first’ company,” and retraining key talent to “think ML” is a central goal. That retraining is no trivial undertaking. According to Christine Robson, who is one of the folks spearheading the initiative, Google invites “folks from around Google to come and spend six months embedded with the machine learning team, sitting right next to a mentor, working on machine learning for six months, doing some project, getting it launched and learning a lot.”
The payoff for one half of a developer-year?
Levy cites comments made by Google CEO Sundar Pichai earlier this year to explain Google’s goals in leveraging ML and the benefits they expect company-wide employees to derive from working with the machine learning team: “Machine learning is a core, transformative way by which we’re rethinking how we’re doing everything. We are thoughtfully applying it across all our products, be it search, ads, YouTube, or Play.”
Models for Remodel
In Mark Edmondson’s introduction to ML, he provides one of the most succinct and helpful definitions of ML:
“Machine Learning gives ability for programs to learn without being explicitly programmed for a particular dataset. They make models from input data to create useful output, commonly predictive analytics.”
Edmondson’s insight is that ML is part of a software engineering thread known as model-driven engineering. ML introduces a new category of model-building activities that can transform the software development life cycle.
Predictive Analytics and ML for z/OS
ML is coming to a mainframe near you, but it may be cloaked in predictive analytics. Last year Zementis, whose products leverage the Predictive Model Markup Language (PMML), announced availability for z/OS. Zementis models can be used to embed predictive models in z/OS CICS or WebSphere settings. The models are “write-once,” meaning they can be deployed to z/OS SPSS, R, Python, or SAS.
In a post on IBM DeveloperWorks, Ravi Kumar outlines how z/OS users can now enable ML on OLTP applications, such as by embedding predictive models in DB2. One technique embeds the z/OS SPSS Scoring Adapter for DB2. Another approach combines a PMML model with business rules to make real-time decisions in DB2 or use Zementis-generated PMML to inject in-app scoring for CICS or Java apps.
The IBM DB2 Analytics Accelerator for z/OS (IDAA) supports several major predictive analytics algorithms: K-Means, Naive Bayes, Decision Tree, Regression Tree, and Two-step.
Remember that saying from Google’s Hal Varian: Google doesn’t have better models — just more data? Varian highlighted volume, but a corollary truism applies to data quality. So far, we’ve only spoken about predictive models, such as for scoring risk or customer retention. Where does ML enter the picture?
Whereas classical hypothesis testing involves a human proposing a cause and effect relationship, ML points software at data to infer cause and effect relationships. As Jeff Erhardt at Wise.io puts it, “Predictive analytics is a use and ML is a technique.”
In Part 2 tomorrow, I’ll explore how IBM is using Predictive Failure Analysis on z/OS, the different flavors for prediction algorithms, how bots can become “development teachers” and how third party tools like Ironstream and Splunk can work together unleash bigger insights from mainframe data.