For people in the know, machine learning is old hat. Even so, it’s set to become the data buzzword of the year — for a rather mundane reason. When things get complex, people expect technology to ‘automagically’ solve the problem. Whether it’s automated financial product consultation or shopping in the supermarket of the future — machine learning is the answer. Data scientists are jumping on the bandwagon, trying to outdo each other in the race for the coolest algorithm. But is algorithm porn bringing us progress, or just a lot of showboating?

SAS Forum Germany in Bonn: how machine learning is used in practice / special offer for Dataconomy readers

Machine learning is no magic bullet. In fact, what’s behind it is basically conventional analytics technology. Analytic models are trained using example data sets. The training is supervised, for instance by specifying the desired output value such as the risk class of a bank customer. The machine is also given the input, in this case master data, demographics, and past transactions. Another example would be providing an error category, with maintenance reports as the input. Non-supervised learning, in contrast, is used to find new patterns in data and learn to distinguish categories.

In other words, the system learns from examples and is able to generalize after the learning phase is completed. What is happening here is not the simple memorization of the examples, but rather the recognition of patterns and laws in the example set. This allows the system to also correctly assess previously unseen instances by transferring what has been learned.

So machine learning helps us develop good models. But a data scientist is still needed to get those models ready for real-world use.

Let’s consider for example the maintenance routine for a CT machine which needs to be optimized to reduce downtime.  First, good models are needed that are capable of taking sensor data and event codes to predict the probability of a component failing to a high degree of accuracy and with minimal false alarms. Machine learning can help here.

The next step is operationalizing, which involves business rules that pair analytic predictions with recommended actions. What should I do if the probability of the motorized patient bed failing is high? How fast do I need to respond if the customer has a premium service agreement? How does the procedure differ if the device is located in a hospital versus a radiology clinic?

The application of the models and the rules must then be continuously monitored. This requires model governance which ensures auditability and the efficiency of the process used to register the models. It also enables automatic accuracy evaluation for of the statistical models and sends out an alert if an analytic model needs to be replaced. And that is something that the data scientist takes care of, not the device technician.

The procedure described above illustrates why machine learning by itself is no magic bullet. In the real world, what counts is the professional integration of analytics into business processes. The goal is not to have the coolest algorithm. Well, at least it’s not the only goal.

At the 2017 SAS Forum Germany on June 29 in Bonn, machine learning will be one of the featured topics — presented from the perspective of both data scientists and the enterprise. Take a look at the program. Use the attractive discount to attend the conference for just €180 (instead of €380). Just select the “Special offer” when you register and then enter the code DCY-SF17 at the end of the registration process.

 

Register now!

 

Like this article? Subscribe to our weekly newsletter to never miss out!

Image: TxDonor

Previous post

A survival guide for the coming AI revolution

Next post

Boost Your Data Wrangling with R