VentureBeat presents: AI Unleashed - An exclusive executive event for enterprise data leaders. Network and learn with industry peers. Learn More
Amazon ships billions of products to customers globally each year, but before those products can be delivered, they have to be packaged carefully to avoid incurring damage-related costs. AI and machine learning play an invaluable part in this — a technical paper describes a system deployed in production at Amazon that automatically assigns packaging types to products, optimizing for package and transportation overhead. The coauthors report that it led to tens of millions in cost savings in emerging markets by reducing the damage rate by 24%.
Were Amazon’s system to be released in open source, it could benefit the countless ecommerce companies faced with the same dilemma, which is likely to become more acute as the pandemic accelerates the adoption of online shopping. According to a Signifyd survey of 10,000 retailers, in-store purchases soared 248% at the end of May compared with a pre-pandemic benchmark.
Retailers like Amazon use a lot of packaging types to ship items from warehouses to customers, which vary in the extent of protection offered during transit. Robust packages provide more protection, resulting in a reduced number of package-related damages, but they’re more expensive due to high material and transportation costs.
For example, Amazon taps polythene bags, custom packs, corrugated T-folder boxes, and carbon boxes (to name a few), all of which come in multiple sizes like small, medium, large, and extra-large. Damages attributable to packaging occur during transit or handling by an associate during the shipment, and they often result in inflated customer compensation and return shipment costs — not to mention adverse effects on customer relationships.
An exclusive invite-only evening of insights and networking, designed for senior enterprise executives overseeing data stacks and strategies.
Amazon’s system takes a two-stage approach to package selection. A model first calculates the probabilities of damage in transit for every product, taking into account the available packaging materials and types along with historical shipment data including product dimensions, weights, fragilities, and the presence of hazardous materials and liquids. To ensure the model remains robust given that very few shipments (less than 0.6%) typically incur packaging-related damages, the historical data is appended with this:
- For every damaged shipment, there are additional shipments with the same product and other inferior (less robust) packaging types, which are considered to be damaged as well. This reflects the notion that if a shipment of a product becomes damaged with a particular packaging type, it’s likely to become damaged in less protective packaging types.
- For every shipment without package-related damages, there are artificially more shipments with the same product and other more protective packaging types that are considered undamaged.
In the second stage of Amazon’s package selection framework, another model considers the tradeoff between choosing protective packaging and settling for an inferior option with a higher probability of in-transit damages. Among other variables, it weighs the packing material cost, transportation costs (including net shipping and total shipping costs computed over all products in an order), package volume, sales velocity of the product over a period of time (except in the case of new products), and the in-transmit damage probability determined by the first model. The end result is a packaging type choice that minimizes shipping costs, ensures products fit snugly, and keeps the potential for damage within an allowable tolerance determined by business constraints.
Before deploying the system into the real world, the researchers fed the “stage one” model shipment data from a three-month period in 2019, which they augmented with the artificially induced inferior and superior packaging types for 100 million examples. They then fed the “stage two” package recommendation model a corpus containing roughly 250,000 products in more than 10 categories with active purchase history in Amazon, with certain business rules like:
- Certain products can’t be shipped due to size or weight, even in the largest containers.
- Liquid and hazardous products aren’t permitted to be shipped in certain countries if they’re not in a certain container.
- “Sensitive” products belonging to certain categories (i.e., adult entertainment) can’t be sent in certain countries without any packaging.
In production, Amazon says that over the course of shipments for 130,000 products, the system decreased transportation costs per shipment by 5%. The “saleability” of products undelivered to customers because of transit damages improved by 3.5%, moreover, with the only negative impact being that the material cost of shipping supplies increased by 2%.
“In many scenarios, the extent of damages depend on the distance shipped, the air or ground mode of transportation used, the quality of the roads along the route, the handling by the courier partners, the location of the warehouses, or even the time of year as during … monsoon season[s], more protection against water or moisture may be needed for some products,” wrote the coauthors. “[P]rotective packaging could be recommended for specific customers who are highly valued or who have had negative delivery experiences in the past. Going forward, we would like to lay emphasis on predicting the optimal packaging type based not only on the product, but using several aforementioned additional factors relating to a specific shipment of an item to a customer. Additionally, we would like to estimate the causal impact of receiving damaged products on customer’s spend patterns and factor it into our optimization algorithm.”
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.