Yandex Lavka Solved the Recommendation Loop Problem with Personalized Exploration

Ramil Boyarchenkov, a machine learning specialist on the Yandex Lavka team, explained on Habr how the service dealt with the problem of recommendation feedback loops — an effect in which a well-trained recommendation system becomes a prisoner of its own accuracy.

The Problem With Good Recommendations

According to the author, the more accurately the system predicts what a user will put in their cart, the less often it shows something unfamiliar — after all, it is more profitable to suggest proven products from the perspective of short-term sales. Over time, the recommendation system becomes locked into already established user habits and stops pushing the user beyond these habits, even when the user's interests change. The system simply doesn't notice these changes because it doesn't see signals beyond the familiar set of products.

An attempt to fix this problem head-on — adding unfamiliar products to the results — almost always hurts metrics: the volume of immediate purchases inevitably declines if you suggest to the user something the system is not confident about.

How the New Mechanism Works

The Yandex Lavka team built a mechanism that adds unfamiliar products not randomly, but personally — to those users who are statistically disposed to try something new.

• Unfamiliar products are added personally, not uniformly to all users
• For each user, an individual probability of showing new items is calculated
• The team calibrated the "aggressiveness" of exploration — the balance between the risk of showing an unsuccessful novelty and the benefit of expanding recommendation horizons
• The author of the material — Ramil Boyarchenkov, machine learning engineer at Yandex Lavka

The idea is to direct exploration to where the potential benefit of introducing the user to a new product is higher than the likely short-term loss in sales — instead of uniformly and therefore more risky adding novelties across the entire audience.

Why This Cannot Be Solved With One Formula

The balance between exploration (trying to show users something new, risking accuracy) and exploitation (relying on already proven preferences) is a classic problem familiar to any recommendation system, from music streaming services to marketplaces. Formally it can be described mathematically, but in practice the correct setting depends on the specifics of a particular business: in grocery delivery, where purchase decisions are made quickly and emotionally, overly aggressive exploration risks disappointing the user with an unfamiliar and unappetizing product, while overly cautious exploration leaves the system locked in already familiar assortment. This is why the Yandex Lavka team had to calibrate the degree of aggressiveness of exploration empirically rather than take a ready-made universal solution.

How This Looks to the End User

From the perspective of a Yandex Lavka shopper, the changes are not directly noticeable: they simply see from time to time in the recommendation selection a product they had not previously purchased and possibly even never searched for. But behind this single display is a calculation of the probability that this particular user is ready to try a novelty without being disappointed with the service. Such targeted personalization of exploration distinguishes the Yandex Lavka approach from more crude schemes where novelties are simply uniformly mixed into the feed for all users in a row, losing some sales on those who do not welcome such experiments.

What This Means

This case is a characteristic example of how a task familiar to any recommendation system — the balance between exploration and exploitation — is solved not by an abstract formula, but by practical calibration for a specific business and specific user. Personalization of exploration allows expanding recommendation horizons without sacrificing short-term sales, and such an approach is applicable far beyond product delivery — in any service where a recommendation system risks becoming locked into products already familiar to the user.