AI predicts before purchase whether a user will return a product

Researchers from Google and Indian e-commerce company Myntra Designs have developed an Artificial Intelligence (AI) that can predict how likely it is that a customer will return a product before the purchase has taken place.

The researchers trained a machine learning model on a dataset of, among other things, preferences, body shapes and viewed products of shopkeepers, writes Venturebeat. In order to identify the factors that had a major impact on the return of a product, the researchers analysed the e-commerce platform of Myntra Designs. There are approximately 600.00 products on it and millions of orders are processed every week.

Training model

The researchers found that of all returned products, 4 percent occurred when there were a number of similar products in a shopping cart. 53 per cent of returned products are due to size and product fit issues. Also, the chance of returning products strongly depends on how much there is in a shopping cart. If it contains five products, the chance is 72 percent, compared to 9 percent of shopping trolleys with one product.

With these insights, the team created a so-called “hybrid dual-model” to predict the chance of returning complete shopping trolleys as well as a few items. A higher level AI classifier classified the returnable trolleys. A second classifier – building on the classified shopping carts – predicts how likely it is that an individual product will be returned.

Both classifiers were trained on a dataset with samples from three categories, with information such as the brand, the age of the product, the size of the shopping cart, the day and time of the order, the city where it should go, the payment method and how often something was ordered.

Results

The best performing AI system was able to predict with 74 percent precision whether an article would be returned or not. In doing so, the system achieved an 82.3 percent range below the receiver operating characteristic, which is a way to determine detection accuracy.

In a live test with 100,000 users, the number of orders decreased slightly (by 1.7 percent) compared to a control set. The percentage of products returned fell by 3 percent.

The team states that it can be useful to determine which customers are likely to return a product, because a retailer can then take action in advance. This can be, for example, the personalization of shipping costs.