A Recommender Dart straight to DIA’s Traditional Coupon Assignment Heart

DIA Group Loyalty Program has been applying personalized promotions over the last 20 years to millions of customers given their individual item consumption. These promotions are based on both items that customers usually buy or don’t buy but should be prone to buy. The goal is to achieve through our customers satisfaction a consistent increase in sales in the medium and long term.

Within this context an improvement area was identified. This was to achieve a higher precision in the search for customers to which we offer certain item or group of items -both by traditional paper coupons as well as digital coupons- of target categories that these customers don’t usually buy and that DIA was willing to promote. On one hand we try to achieve an increase in coupon redemption and on the other obtain a higher Return Of Investment from these promotional coupons.

At the same time, under a more traditional view, we also wanted to offer to each customer those items and item categories that he or she is not actually buying but is prone to buy. The achievement of a higher precision in this kind of personalized assignment promotes the development of the customer’s basket and makes the up-selling strategy more efficient.

The great challenge here was to compete with an already established tool and human know-how of more than 20 years of experience that was attaining good redemption results in this kind of coupons. Another challenge was to manage a huge data volume and being capable of integrating a new development, with the existing Loyalty Department coupon assignment application, in an agile manner. This is an on-premise application that serves more than three thousand million coupons to DIA’s customers in one year.

Under this ecosystem a multidisciplinary team composed by DIA’s Big Data Client Product Owner, DIA’s Big Data Department Senior Data Scientist, Big Data Engineer and a BBVA Next Technologies team by a Scrum Master, Cloud Architect, Dev Ops, Software Developer and Data Scientists have been working during several months on developing and productionizing new DIA’s Coupon Recommender System. We can proudly say that after a few months on production we are obtaining very positive online results.

The implemented model is based on collaborative filtering techniques that learn from implicit data, defined as the boughts performed by DIA’s customers. More in detail, these techniques learn latent knowledge in the data, characterizing both items and customers under a priori not interpretable factors. These latent factors are assumed to capture certain dimensions like how healthy an item and/or customer are, for example. They also enclose information such as buying habits patterns, lifestyle, family size and customer age range, although none of this information is given as an input to the model. As a result the model learns where to place each item and customer over each factor, being represented by a numerical value that tells how likely a customer-item match is to occur.

Finally, it is worth mentioning that we have not only achieved our initial goals, but surpassed them over our own expectations. This has lead to an invaluable knowledge acquisition and has opened new possible applications that were not even considered at first. As an example the learned latent factors can serve as source for characterizing and clustering DIA’s assortment in a different way, based on buying habits, and can help to validate and rethink already applied models in the company developed for assigning coupons with items bought by our customers.