The Secret to Success: How Recommender Systems Are Changing Industry
A scientific conference dedicated to the young and actively developing field of recommender systems was held at HSE University. Representatives of the scientific community and industry gathered at the site to exchange cutting-edge ideas and best practices, as well as discuss opportunities to implement new technologies in real-life business scenarios.
The Research and Educational Laboratory for Matrix and Tensor Methods in Machine Learning at the Institute of Artificial Intelligence and Digital Sciences at HSE University’s Faculty of Computer Science organised the conference ‘Recommender Systems in Research and Industrial Applications.’ Scientists from HSE University, Sberbank, AIRI, MTS, Tinkoff, and OK presented reports on current topics in this area, including validation algorithms, user action predictions, usage of large language models, etc.
Recommender systems are software tools that offer users personalised recommendations and suggestions based on their preferences, profile, activity history, and other data. They are used in various fields, including e-commerce, streaming platforms, social networks, news portals, etc. Recommender systems play an important role in business development. They help increase sales and profitability by offering relevant products and services to users, which leads to higher conversion and increased revenue.
Maxim Rakhuba, Head of the Research and Educational Laboratory for Matrix and Tensor Methods at the Faculty of Computer Science
‘The topic of recommender systems is related to what our laboratory does. For example, some widely used recommender system models are based on matrix factorisation. However, it is quite easy for a person working in academia to fall into the trap of developing some beautiful and reasonable method of matrix factorisation, which may, for various reasons, not be suitable for industrial applications (or may well be in its rightful place). Therefore, there is a need for a dialogue between academia and industry, which is what we tried to achieve at the conference. In addition, many interesting scientific and engineering approaches are now actively developing. And we wanted to exchange new developments, experiences, and best practices at the conference.’
Evgeny Frolov, Associate Professor at the Faculty of Computer Science and Senior Research Fellow at AIRI
‘Recommender systems are one of those areas where the industry largely guides scientific discourse and shapes the demand for innovation. Historically, the composition of scientific conferences on recommender systems has been approximately equally divided between representatives of the scientific community and high-tech companies. This leads to an intensive and fruitful exchange of advanced ideas and developments, the creation of new fronts of research, and the delineation of problem areas in understanding the tasks and approaches to solving them. However, until now, the practice of conducting such events included mainly other countries, while Russia has a large, advanced community of experts involved in various aspects of the development and implementation of recommender systems. In this regard, the creation of a local platform for dialogue is a significant event, which, I hope, will gain momentum and become regular.’
Daria Tikhonovich, MTS
‘Classical approaches to validating recommendation models can have serious problems when it comes to choosing an algorithm for a real service. In practice, popularity bias (when a recommendation system shows users the most popular object—Ed.) is almost always present in the data. As a consequence, validation using ranking metrics will result in the selection of a model that is biased in recommending popular content. In my report, I show alternative approaches to validation that better reflect the usefulness of recommendation models for users and businesses and help build relevant, personalised recommender systems. The approaches proposed in the report are now being developed as open-source tools for advanced model validation in the RecTools library.’
Nikita Bezlepkin, Sber
‘The report described current trends in recommender systems based on the results of the RecSys 2023 conference in Singapore. One of the trends for 2023 is the use of large language models, which can improve the performance of recommender systems. Cases from companies such as Amazon, Netflix, ByteDance, etc. were reviewed. Real cases of using LLM (large language models—Ed.) in Sber ecosystem products were also presented.’
Sergey Ermilov, ОК
‘Matrix factorisations remain a popular choice for developing modern recommender systems. They can also be successfully generalised to the multidimensional case using tensor expansions. The application of such methods in recommendations allows for the use of additional information sources and the creation of more accurate recommendations. Research on tensor approaches is being conducted in academia, but their application in industry is limited due to the huge volumes of data. We have implemented tensor decompositions in a distributed environment that allows these methods to be used on real-world problems and bridge the gap between the academic and industrial applications of similar methods. In the future, we will conduct experiments on our own tasks to evaluate the effectiveness and practical applicability of tensor approaches in recommendations.’
The round table at the conference was devoted to the development and implementation of recommender systems in company products, the redistribution of personnel, and cooperation between industry and universities for solving complex and long-term tasks. The conference participants noted the existing gap between science and industry.