T-Bank AI Research: Sampled Maximal Marginal Relevance (SMMR) Method of Accelerating Recommendation Algorithms

The main articles are:

• Big Data
• Data mining Data mining

2025: Creating a method to accelerate recommendation algorithms

On July 14, 2025, T-Bank AI Research announced the creation of the Sampled Maximal Marginal Relevance (SMMR) method, which allows up to 10% faster and more diverse recommendation formation on online platforms compared to other methods known in science. The method helps to create more personalized collections for the user's interests, but not focusing on one type of content.

𝓲

Фото: T-Bank

According to the company, it T-Bank plans, first of all, to introduce this method into its own digital services to optimize the quality of recommendations - in T-Shopping - to form wider and more flexible product sets, and in - social network Pulse to increase diversity in the user feed.

Traditional algorithms seek to select the most suitable objects - goods, films, news - based on the user's preferences. However, this approach often leads to recommendations of the same type, forming the so-called "information bubble," when the user sees only those goods or content that are similar to his previous interests. For example, if the user often watches comedies, personalization models will only show this genre without offering alternatives.

A similar problem is observed in news feeds: the algorithm fixes clicks on the most noticeable headlines and begins to promote similar content, even if the material itself is not interesting to the user or his needs for information are much wider.

The SMMR method solves this problem through probabilistic choice: instead of giving preference to one most relevant object each time, the algorithm randomly selects from a limited range of suitable options. This approach provides a greater variety of content and helps the user discover new knowledge.

In addition to increasing diversity, SMMR is also faster than its known analogies - MMR (Maximal Marginal Relevance) and DPP (Determinantal Point Process). The algorithm selects multiple objects in one iteration and optimizes the sample size with each step, reducing the number of required steps from 100 to 5-10 on the sample size of 3,000 candidates. As a result, this technique turned out to be 2-10 times faster than MMR and DPP and at the same time ensured an increase in the variety of recommendations by 5-10%. The SMMR method randomly selects similar categories for display.

The SMMR method represents the development of classical approaches to the formation of various recommendations and is distinguished by a number of fundamental features:

• Probabilistic approach to selection. Unlike traditional methods, which select the most relevant object at each step, SMMR uses a probability sample from a limited number of candidates. This allows for increased diversity without significant loss of relevance.
• Batch sample with increasing size. The algorithm generates recommendations not for one, but for several elements, gradually increasing their size. This approach allows you to optimize the operation of the model, especially on large samples, by reducing the number of iterations.
• Controlled level of randomness. The method includes the "temperature" parameter, which controls the degree of randomness when selecting objects. This makes it possible to customize the algorithm for specific tasks - from conservative selection to more diversity in issuance.
• Scalability on big data. By reducing computational complexity, SMMR works efficiently on large amounts of data. For lists of several thousand objects, the algorithm requires 10-100 times fewer iterations than classical approaches.
• Easy integration into your existing infrastructure. The method is easily integrated into current recommendation systems. This reduces implementation costs and optimizes the technology adaptation for business tasks.

Recommendation systems perform a much more complex task than just guessing the interests of the user. Their goal is to offer content that a person might not have discovered on their own. For example, when searching for books in an online store, traditional algorithms usually recommend 10 works of the same genre, based on user preferences. Unlike them, SMMR technology complements the selection with 2-3 books from other genres - popular science publications or thrillers - which may interest the user, but previously did not fall into the field of view of the algorithm. This approach helps to find the optimal balance between the accuracy of recommendations and the element of surprise, which contributes to the involvement and expansion of user experience.

The application of the SMMR method can be of practical benefit to both business and users. For digital platforms, this means an increase in audience satisfaction: recommendations become not only relevant, but also diverse, which helps to retain attention and optimizes engagement. For example, instead of ten films of the same type, the streaming service can now offer five predictable and five unexpected, but potentially interesting options. It also contributes to a more even coverage of the catalog - more pieces of content get a chance to be shown, which is beneficial for sellers, manufacturers and aggregators. For users, this opens up various opportunities: instead of focusing on one genre or commodity category, they receive a variety of offers, which makes interaction with the service more meaningful and personalized.

The SMMR method has been tested on three open datacets: MovieLens (movies), Dunnhumby (purchases) and MIND (news). It demonstrates stable results both in consumer scenarios (selection of films, goods) and in more dynamic ones - for example, in news recommendations.