Фролов Евгений Петрович

Проекты

• Successfully negotiated and delivered several projects with large companies: secured time, budget, and scope
  of work; organized and managed diverse teams of researchers of up to 7 people. Details:
  Joint projects with Huawei (Principal Investigator).
  ◊ Hyperbolic Geometry for Sequential Learning algorithms in next item recommendations. 2022 – 2023.
  ✓ Lead the development of several architectures for sequential learning with hyperbolic layers.
  ✓ Designed a new solution combining hyperbolic classifier with sequential self-attention.
  ✓ Established connection between data properties and compatibility with hyperbolic models.
  ◊ Privacy-preserving Federated Learning for next app predictions on mobile platforms. 2021 – 2022.
  ✓ Lead a cross-disciplinary team. Ensured all components of the designed solution are compatible
  with practical requirements of the customer. Developed a novel matrix factorization approach.
  ✓ The final solution was accepted and tested in pre-production environment by the customer.
  ◊ On-device user actions prediction for quick search of relevant apps. 2019 – 2020.
  ✓ Lead a team of researchers from different universities. Developed several new solutions.
  ✓ Provided solution satisfied the target accuracy specs and was implemented in production.
  Collaboration with Sber AI Lab. 2018 – 2020.
  ◊ Joint project on Hyperbolic Autoencoders for recommender systems.
  ✓ Initiated and delivered a cutting-edge research project related to application of hyperbolic
  geometry in collaborative filtering models. Presented results at the main RecSys conference.

  ◊ Consulting:
  ✓ Consulted the design of “Stories” recommendation in the bank’s mobile application. A/B testing in production indicated a 56% increase of the conversion rate.
  ✓ Implemented a prototype restaurant recommendation system based on user transactions. Part of the solution was used in the production system.
  Other related activities:
  ◊ Collaboration with R&D Lab of prudSys company (https://prudsys.de), 2014 – 2017.
  Conducted research on application of tensor methods for online recommendations in retail.
  ◊ Open-source
  ✓ Developed Polara, an open-source recommender systems framework. The framework is used by individuals as well as several companies for quick prototyping and as a tool for performing exploratory research. Repository: https://github.com/evfro/polara
  ✓ Developed fast PyTorch loaders for collaborative filtering data: https://github.com/evfro/cofida

   

Гранты

Опыт работы