We are excited to bring Transform 2022 back in-person July 19 and virtually July 20 - 28. Join AI and data leaders for insightful talks and exciting networking opportunities. Register today!
With sales anticipated to reach $4 trillion by 2020, ecommerce is an unstoppable force. It’s no real mystery why — customers prefer the convenience of online stores compared with brick and mortar, countless surveys show — but it’s a mixed bag for companies that attempt to edge out rivals with liberal return policies. It’s estimated that as high as one-third of online orders are returned, which whittles down profit margins.
Researchers at Google and India-based ecommerce company Myntra Designs took a stab at the returns pain point in a new paper published on the preprint server Arxiv.org. Using a machine learning model trained on a data set of shoppers’ preferences, body shapes, product views, and more, they managed to predict per-customer return probability prior to purchase.
To identify which factors had a disproportionate effect on returns, the researchers conducted analyses on Myntra Designs’ ecommerce platform, which hosts approximately 600,000 products at any one time and facilitates millions of orders a week. They found that out of all returns, 4% occur when there are some similar products present in the cart. Additionally, they discovered that 53% of returns are due to size- and fit-related issues and that return rates are highly dependent on cart size, with cart sizes exceeding five products seeing return rates around 72% compared with 9% for carts with one product. And perhaps unsurprisingly, older inventory has an almost doubled return rate compared to newer items.
Armed with these insights, the team devised what they call a hybrid dual-model to forecast both cart and item return probability. A higher-level AI classifier classified returnable carts, while a second classifier (drawing on the carts classified as returnable by the first classifier) predicted return probability at an individual product level. Both trained on a data set containing samples across three categories — product-, cart-, and user-level features — including (but not limited to) things like brand, product age, cart size, order day and time, delivery city, order count, payment mode, and purchase frequency.
So how’d it perform? In experiments, the best-performing return-predicting AI system achieved 83.2% area under the receiver operating characteristic (AUC) — a measure of detection accuracy — and 74% precision. In a live test conducted with 100,000 users, the number of orders dipped slightly (by 1.7%) compared with a control set, but the return percentage dropped by 3%.
The team notes that knowing which customers are likely to return an item enables a retailer to take preemptive actions, like personalizing shipping charges or making the product no-returnable by offering a coupon. “As future work, we plan to apply this model on more action items which can further help in reducing the overall returns,” they wrote.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn more about membership.