Researchers at the University of California (UC Riverside), in the USA, developed algorithms capable of recommending products based on consumers’ shopping habits. Using a methodology called tensor decomposition — used to find patterns in large volumes of data — the system makes suggestions tailored to the customer’s preferences.
These tensors are represented as multidimensional cubes, used to model and analyze information with many different components. Data closely related to other common factors can be connected to discover patterns that are not noticeable in a first layer of observation.
“Each tensor has three different modes to capture an aspect of that transaction. Consumers form one mode, while the second and third capture product-to-product interactions, considering everything that has been purchased before,” explains computer science professor Negin Entezar, lead author of the study.
Want to stay on top of the best tech news of the day? Access and subscribe to our new youtube channel, Kenyannews News. Every day a summary of the main news from the tech world for you!
Focus between the lines
The researchers used a hypothetical situation, analyzing the buying patterns of three different consumers as proof of concept. In a single transaction, customer A purchased sausage, bread, soda, and mustard. Customer B took the same products, but in separate purchases. And customer C didn’t buy soda.
For a conventional matrix-based algorithm, customer A is identical to customer B in that they buy the same items. However, when using tensor decomposition, client A is more closely related to client C because the behavior of the two is more similar. Both purchased similar products in a single transaction, although their purchases were slightly different.
“The typical recommendation algorithm makes predictions based on the item the customer has just purchased, while the tensor decomposition makes suggestions taking into account all products in the bag. If a shopper has dog food and peanut butter in their basket, but no bread, the algorithm will suggest a chew toy for their pet instead of jelly.”
data modeling
Tensors are multidimensional structures that allow the modeling of complex and heterogeneous data. Instead of just noticing which products are purchased together, they look at a third dimension and make suggestions based on the habits of other consumers who have carried out similar transactions.
With this approach, it is possible to expand the scenario of complementary product recommendations made through specialized algorithms, making the item suggestions much more accurate and faithful to the purchase behavior of specific consumers.
“Tensor methods work because they look at the whole and not just one recurring factor. Although they are very powerful tools, they are still more popular in academic research and, for the industry to adopt them, we must show that it is possible to painlessly replace the recommendation systems used today”, concludes Papalexakis.
