Cargo’s Crystal Ball: The power of machine learning

It all starts with data

Despite being classified as “machine learning,” the process of developing a functional algorithm starts with a team of very human data scientists who first determine the type of solutions, or “predictions,” an algorithm should ultimately be able to make on its own, and then teach the algorithm “how” to learn independently, without the need for human guidance.

When researchers within the logistics division of JD.com, one of China’s top e-tailers, set out to boost warehouse efficiency through machine learning, the process of optimizing inventory placement began with data regarding past customer purchases. With this data, an algorithm was taught to identify relationships between products, and from this data “suggest combining them in the adjacent areas or in nearby shelves, improving efficiency when warehouse employees pick up goods,” said Cheng Yan, head of research and development for JD Logistics.

Although a lack of data is a common complaint, JD.com says it has plenty. The e-commerce giant, much like its American counterpart Amazon, relies heavily on its own data to improve processes. Not only does JD collect information on its 236.5 million customers, “We also do much of our domestic logistics in-house, including long-haul trucking and last-mile delivery. This makes it easy to both collect data, and to implement our machine-learning technology.” JD’s supply chain partners also supply the e-tailer with data troves, said Yan.

Access to sufficient data alone, however, is useless if the data is unclean or not properly categorized. “Within the vast quantities of data we collect from every part of our supply chain, a small amount is created with human input, which can introduce some variance,” said Yan.

Improving the algorithm

Supply chain management company elementum is one firm that is aggressively working to isolate relevant data in order to produce better insights. “It’s not about capturing all the data, it’s about putting it together in a way where it can be consolidated, categorized and actionable,” said Gregg Brody, head of carrier success.

For the Silicon Valley offshoot of multinational technological manufacturer Flextronics, machine-learning capabilities are enhanced when subject matter expertise and talented engineers coalesce. Many people at tech firms unfamiliar with air cargo wrongly assume that airfreight is bought, sold, and moved in a manner analogous to ticket-purchasing passengers. When it comes to building and improving the logic that guides machine-learning algorithms applied to the air cargo industry, it is important to understand how freight actually moves.

In order to teach machine-learning algorithms how to interpret the finer nuances of air-freight movements – from zone-skipping to the many iterations of consolidations – elementum has developed a system to field and properly categorize outliers that do not reflect traditional freight forwarding brokerage functions. “What we’ve designed is a program where there is a liaison between when the exceptional information is captured, and, if it is different from an existing category, we then put engineering together with industry experts to develop a new category.” By improving the underlying logic behind the interpretation of available data, Brody said, machine learning can utilize the data to create better insights for scenarios that are relevant to the end user.