RocketHuntSmart job search
Reach out directly about this role
ML Developer for Yandex Food
Yandex Food is a rapidly developing service providing food delivery from restaurants and shops to millions of users across more than 10 countries on different continents.
Our team manages the algorithmic and ML components for ranking in Yandex Food's customer product. We develop personalized recommendations, improve search, evolve new user product scenarios, and also help launch ranking for new countries. We actively implement both classic machine learning methods and advanced neural network developments to enhance ranking quality in Yandex Food.
More about what we do: ML in Yandex Food: how it works and who is involved?
There are currently six ML developers in the recommendation algorithms group, but we are expanding and looking for new people to join the team.
Our internal collaboration processes are well-established with minimal bureaucracy. We solve complex problems at the intersection of product development and algorithms; our implementations directly impact Yandex Food's business and product.
Examples of projects you will work on:
1. Transition to end-to-end ranking and CPA auction in advertising
In Yandex Food, the ad auction worked on a CPC model (cost per click — a partner paid for a click on their restaurant when it was raised to the top of search results due to advertising) for a long time. Four slots at the very top of the search results were allocated to ads, and the higher the partner's bid, the higher the probability of making it to the top.
Very recently, we transitioned to a CPA auction (cost per action — where a partner pays for an order if their restaurant was raised in the search results by advertising). We created a mechanism where organic search results are mixed with ads, and through bidding, partners have the opportunity to rise above organic restaurants. To achieve this, we had to move from predicting relevance to predicting the probability of an order and its margin, and also use simulations to select optimal mixing coefficients for ads and organic results. In the future, we plan to apply Bayesian optimization for more optimal selection of mixing coefficients, switch to other ranking targets, and much more interesting work.
2. Launching ML ranking in "Where to go"
"Where to go" is a new product within Yandex Food that allows users to choose a restaurant to visit in person if delivery has become tiresome. For the launch, we needed to solve the cold-start problem. For this, based on data from Yandex Crypta, we prepared a DSSM model capable of generating LaL vectors for users and restaurants. We also did extensive infrastructure work preparing the logic for feature calculation and logging, in order to later transition to the target scheme — CatBoost ranking.
Currently, in "Where to go," we are implementing collaborative models SLIM and iALS. In the future, we plan to extract value from new data sources, switch to new targets, and also improve search ranking.
3. Improving upsell in retail
Upsell is a recommendation feed with products on the cart page for shops. We've done extensive work to improve recommendation algorithms on this surface: we implemented mixigen (a model that determines which sources to recruit candidates for ranking from), added many personalization factors, and implemented userbody (a large two-tower model created within the depths of Yandex's Big Search). In the future, we plan to personalize mixigen, work on improving product mechanics, and add new candidate sources.
Learn about the development of Yandex city services at dev.go.yandex
Responsibility for all stages of ML projects You will participate in designing machine learning-based solutions to improve ranking, help translate business requirements into ML tasks, and develop and validate the correctness of ML pipelines.
Product improvement and hypothesis testing in A/B experiments You will search for growth points in the product, propose and defend ideas for its improvement using ML ranking. You will test your ideas during A/B experiments, analyze them, and defend the results.
Writing production code Our models operate in real-time services processing hundreds of requests per second. You will write production code in C++, test it, and optimize it if necessary.
Interaction with related teams Great things are not done alone, so you will work in a large cross-functional team that includes analysts, developers, ML specialists, and managers. You will closely interact with related teams for joint progress.
3-5 years
Experience
Full-time
Employment
Hybrid, Onsite
Work Format
Middle
Grade
Data Science & ML
Specialization
IT & Tech
Industry
Corporation
Company Type
By city
Data Science & ML
Specialization
IT & Tech
Industry
Corporation
Company Type