machine learningChief supply chain officers plan to invest in analytics, business intelligence, and other software tools to bring more visibility to their supply chains, according to “Orchestrating a customer-activated supply chain,” a report issued by the IBM Institute for Business Value. IBM canvassed 201 chief supply chain officers from 16 countries for the research.

The study found that 92 percent of supply chain executives said they expect to have implemented advanced analytics in the next two to five years, and 80 percent said the same for modeling to optimize flows. (Those figures include a small number of respondents who already have those capabilities in place.) In addition, 62 percent said that in the next three to five years they plan to invest in software and tools that will help them achieve supply chain visibility.

Analytics is of course a very wide area, we would like to focus on one technology that has not been implemented widely in supply chain until recently called Machine Learning and in particular how it can be combined with optimization to produce breakthrough results.

A good definition of machine learning is here: “Machine learning is about learning to do better in the future based on what was experienced in the past. The emphasis of machine learning is on automatic methods. The goal is to devise learning algorithms that do the learning automatically without human intervention or assistance. The machine learning paradigm can be viewed as programming by example.  Often we have a specific task in mind, such as spam filtering. But rather than program the computer to solve the task directly, in machine learning, we seek methods by which the computer will come up with its own program based on examples that we provide.”

Some examples of machine learning include demand and price forecasting, character or face recognition, medical diagnosis, fraud detection, topic spotting (such as trending news).

There are several types of machine learning results – the main ones are:

Regression – this is where the values are real. For instance if you need to price a product or when putting your house on the market – the machine learning algorithm will learn from previous sales and predict what is the product price based on this information.

Classification – this is where the values are discrete – for instance whether you will be accepted into a certain program or whether you have a certain disease or not. The machine learning algorithm will again look at previous cases and predict whether your characteristics place you in or out of certain groups.

Another aspect of machine learning is what type of process is used – these can be:

Supervised – this is where the information includes the result such as providing a list of characteristics and the resulting price or conclusion and using these to predict a new case. This is typically used in price forecasting but also used in recommendation engines as well as in character and face recognition.

Unsupervised – this is where information is provided with no clear conclusion in mind and the algorithms come up with correlations based solely on the data. This is used for data mining and has become more sophisticated in recent years with the use of visualization to help data matter experts  determine the meaning of the results. This type of process has applications in fraud detection, genetics analysis and finance.

A natural application of supervised machine learning in supply chain analytics is forecasting. This is true in particular for fashion products where there is low volume, no similar products and high volatility. Another application is in customer segmentation where grouping of different customer characteristics can be enhanced with machine learning techniques.

Unsupervised machine learning has been used to scour customer records for unexpected correlations in buying habits and can also be used in the same way to look at supplier information for new clues for determining quality and risk.

While a machine learning approach enhances the quality of grouping and forecasting, the results are often independent of each other. In order for a system to work in a meaningful way, it needs to make sure all the parts are optimized relative to each other.  This is exemplified in recent work that David Simchi-Levi and his team at MIT did with Rue La La.

Rue La La is an online fashion sales company offering extremely limited-time discounts (“flash sales”) on designer apparel and accessories, see an overview of the project.

One of Rue La La’s main challenges is pricing and predicting demand for these first exposure styles which was reflected in either quick sale-outs or too much leftover inventory. Therefore, Rue La La first approached MIT to reduce inventory in their supply chain. The team started looking at historical data and discovered that by setting prices using historical data it could solve both problems. This approach involved using a combination of machine learning in order to predict demand for new items and estimate lost sales followed by optimization in order to take into account competing styles when setting prices.

This approach produced increased revenues of 10% and won the team the 2014 INFORMS Revenue Management and Pricing Section Practice Award.

At OPS Rules, we are using the same method in several different customer applications. We believe that this combination is a winning approach to implementing the power of machine learning technology to operations problems.

Written by Edith Simchi-Levi, VP of Operations at OPS Rules