A Tournament of Machine Learning Champions

Behind today’s powerful predictive applications are machine learning models that identify patterns in the data. But getting these models trained and tuned is not an easy process. Analytics giant SAS thinks it has a solution with a new offering called Factory Miner that allows you to run tournaments to pick the best machine learning model for a particular data set.

Factory Miner lets users build and test a slew of different machine learning models using seven different algorithm families. Testing more models and algorithms not only increases the odds that the user will be able to identify the best one for a particular use, but it also allows users to get more fine-grained with their predictive approaches, says Sascha Schubert, director of analytics product marketing at SAS.

“What we do with Factory Miner is we provide an automated environment to build lots of models quickly, especially for things that we call segmented or stratified modeling,” he tells Datanami in an interview today. “Many organizations are going to smaller and smaller segments, but more and more of them. What we see in our engagement is that the models get more accurate when you can specify your segments more precisely and build models for these particular segments, rather than build a model across the entire base.”

Factory Miner automates the entire model-building process. That includes everything from data ingest and exploration to generating models and running model tournaments. The software even automates the process of re-training the models and auto-generates the code that can be deployed in SAS Analytics, Hadoop, Teradata, Greenplum, and IBM Netezza database environments.

Factory Miner ships pre-loaded with seven popular machine learning algorithms, including: Regression, neural networks, decision trees, random forest, Bayesian networks, support vector machines, and gradient boosting. The tournaments pit different models against each other against a given data set, and the software supplies the statistics to tracks which ones worked best.

Who’s in your machine learning Final Four?

“You find that different models work differently in different segments,” Schubert says. “It’s good to have different modeling techniques available, and Factory Miner allows you to run these models in an easy-to-use and simple way to find the best performing model quickly.”

Factory Miner enables less technical users to test hundreds of machine learning models against their data. The software, which features a Web-based UI, lets users tweak the models if they choose to, although it’s not needed, and also lets users create a “best practices” template that they can run in the tourneys as a sort of baseline.

All machine learning models need to be re-trained periodically, and Factory Miner also automates this to ensure that it’s generating accurate results. “If your model is in production for six months or a year, the causal relationships that are used to build that model might change, so your model might not perform as well anymore,” Schubert says. “When you have new data coming in one, two, or six months later, you can automatically retrain it.”

The shortage of data scientists and the need to iterate quickly is pushing predictive analytics towards more software automation. Today, automated generation of machine learning models is somewhat of a rarity, but it’s becoming more popular all the time. Predictive analytic software outfits like Skytree and H2O.ai are increasingly looking to speed the process of generating and training predictive models as part of a shift towards predictive model factories.

SAS Factory Miner lets users compare machine learning models

SAS is also looking to capitalize on this trend and give its customers access to faster and more pertinent insight from their investments in predictive analytics. Up to this point, SAS users would need to hand-code each parameter of their machine learning models in Enterprise Miner, the company’s powerful but also complex data mining product.

With Factory Miner, SAS is helping mere data analysts do some of the work that previously required data scientists–and it’s helping them to do it faster to boot. “We want to make these data scientists more productive and help the organization to expand their talent pool, and democratize predictive analytics,” Schubert says.

Related Items:

How Machine Learning Is Eating the Software World

The Rise of Predictive Modeling Factories

Inside Cisco’s Machine Learning Model Factory