Yandex Lavka breaks the recommendation feedback loop with personalized exploration

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

What's wrong with good recommendations

According to the author, the better a system gets at guessing what a user will put in their cart, the less often it shows anything unfamiliar — pushing proven products pays off more in short-term sales. Over time, the recommender locks onto a person's established habits and stops nudging them beyond those habits, even when the user's interests change. The system simply misses these shifts because it sees no signals outside the familiar set of products.

The brute-force fix — adding unfamiliar products to the results — almost always hurts metrics: near-term purchase volume inevitably drops when the system offers users items it isn't confident about.

How the new mechanism works

The Yandex Lavka team built a mechanism that mixes in unfamiliar products not at random but on a personalized basis — targeting users who are statistically inclined to try something new.

• Unfamiliar products are mixed in per user, not uniformly for everyone
• Each user gets an individually calculated probability of being shown new items
• The team calibrated the "aggressiveness" of exploration — the balance between the risk of showing a bad new item and the benefit of broadening the recommendations
• The article's author is Ramil Boyarchenkov, a machine learning engineer at Yandex Lavka

The idea is to steer exploration toward cases where the potential payoff from introducing a user to a new product outweighs the likely short-term loss in sales — rather than spreading new items evenly, and therefore more riskily, across the entire audience.

Why a single formula won't solve this

Balancing exploration (trying to show the user something new at the cost of accuracy) against exploitation (relying on already proven preferences) is a classic problem familiar to every recommender system, from music streaming services to marketplaces. It can be described formally in mathematical terms, but in practice the right tuning depends on the specifics of a particular business. In grocery delivery, where purchase decisions are fast and emotional, overly aggressive exploration risks disappointing users with an unfamiliar product they don't like, while overly cautious exploration leaves the system stuck in the familiar assortment. That's why the Yandex Lavka team had to calibrate exploration aggressiveness empirically instead of adopting an off-the-shelf universal solution.

What the end user sees

From a Yandex Lavka customer's perspective, the changes aren't directly noticeable: they simply see, from time to time, a product in their recommendations that they've never bought before and perhaps never even searched for. But behind that single impression is a calculation of how likely this particular user is to try a new item without becoming disappointed with the service. This targeted personalization of exploration sets Yandex Lavka's approach apart from cruder schemes that simply mix new items evenly into everyone's feed, losing some sales among users who don't welcome such experiments.

What it means

This case is a typical example of how a problem familiar to every recommender system — the exploration-exploitation balance — gets solved not with an abstract formula but with practical calibration for a specific business and a specific user. Personalizing exploration makes it possible to broaden recommendations without sacrificing short-term sales, and the approach applies well beyond grocery delivery — to any service where a recommender risks getting stuck on products the user already knows.