Learning not to over-think predictive analytics
The predictive analytics and machine learning markets are projected to grow at a rate of 15 percent annually through 2021, yet many organizations fail to reap full benefits from their investments. The problem is often that the organization makes the process too complicated.
The irony is that most large organizations are more prepared to implement and use predictive analytics and machine learning than they think, says Mike Gualtieri, a research analyst with Forrester Research.
Gualtieri has just published a Forrester Wave report on predictive analytics and machine learning, “The Forrester Wave: Predictive Analytics and Machine Learning Solutions, Q1 2017.” In it, Gualtieri notes that organizations that want to leverage artificial intelligence need to start with a predictive analytics and machine learning solution.
“Predictive models created using machine learning are already commonly used for marketing, customer intelligence, and risk models,” Gualtieri says. “The teams of data scientists that create these models are in the know. But, often the employees such as enterprise architects who are charged with investigating AI don't understand that machine learning models are fundamental building blocks of AI.”
That is unfortunate, Gualtieri says. “There are hundreds, if not thousands, of opportunities to use machine learning models in business processes and customer experiences. This is not day one, but it is still only day 2. There is tremendous opportunity today, but most enterprises struggle about how to think about AI. They are thinking too big. Successful machine learning models is about predicting one simple thing that can have a big impact on the business such as the next best product to recommend for an individual customer.”
Forrester forecasts a 15 percent compound annual growth rate (CAGR) for the PAML market through 2021. That’s a conservative estimate, given that the PAML category includes and overlaps with AI and deep learning. Gualtieri says the category continues to be hot since most large enterprises want the power to predict and have only scratched the surface of what is possible.
“The best PAML solutions have model management,” Gualtieri notes. “The dirty little secret about machine model is that they are based on correlations that work on historical data. That can be a problem since the models created using that historical data are used to predict future outcomes.”
“Mature organizations and data scientists have a healthy ‘the model works until it doesn't work’ attitude,” Gualtieri continues. “Model management is a feature of PAML solutions that monitors the effectiveness of those models in real-time to make sure they are still effective. Mature vendors such as FICO, SAS, and IBM understand this better than anyone. The startups have work to do.”
The 15 percent Forrester growth prediction for PAML doesn't include deep learning, Gualtieri says.
“I predict that in 2017 deep learning will have a 700% increase in adoption - but that adoption by large enterprises will be for experimentation,” Gualtieri says. “All of these PAML vendors see it coming and most have a strategy for including deep learning algorithms in their software.”
“Having said that, the open source community is wonderfully active in creating new algorithms accessible in popular programming languages such as Python, Apache Spark, and tools for data scientists such as Juypter,” Gualtieri says. “The PAML vendors are on top of this. Startup Dataiku has a very promising mix of the open source innovation and enterprise needs.”
Finally, Gualtieri says “The number one mistake of the PAML startups is that they build a product that is attractive to data scientists only. They must also build a product that is attractive to chief data scientists.”