RU-2024132896-A - SYSTEM AND METHOD FOR FORMING RECOMMENDATIONS OF DIGITAL ELEMENTS

Inventors

• Бурлаков Даниил Сергеевич
• Карпова Анастасия Евгеньевна
• Сафронов Александр Валерьевич

Assignees

• ОБЩЕСТВО С ОГРАНИЧЕННОЙ ОТВЕТСТВЕННОСТЬЮ "ЯНДЕКС"

Dates

Publication Date

20260504

Application Date

20241101

Claims (20)

1. 1. A computer method for generating recommendations of elements for users of a digital recommendation platform storing a plurality of digital elements, which includes:
2. - training a machine learning algorithm to find previously undisplayed elements associated with sources previously unknown to a given user, in order to recommend them to that user, including:
3. - obtaining for a given digital element from a set of digital elements a vector of an element that characterizes the features of this digital element, including the source associated with it;
4. - obtaining for a given user of a digital recommendation platform a user vector characterizing the features of this user;
5. - finding in the user's log associated with a given user a training set of past digital elements from a plurality of digital elements of a digital recommendation platform, wherein the given past digital element from the training set of past digital elements is associated with an indication of the interaction of this user with this element;
6. - the formation of a training data set containing a plurality of training digital objects, wherein the formation of a given training digital object from the plurality of training digital objects includes:
7. - scanning a training set of past digital items arranged in chronological order to find, for a given past time, a given target past digital item from a target source that is the first digital item from among the digital items associated with the target source that a given user interacted with prior to the given past time;
8. - forming a training subset of past digital items containing past digital items with which a given user interacted before interacting with a given target past digital item;
9. - formation of a given training digital object containing a user vector, a combination of element vectors from a training subset of past digital elements associated with a given target past digital element, and a label containing an indication of the interaction of a given user with a given target past digital element associated with a target source;
10. - transferring each of the plurality of training digital objects to a machine learning algorithm, as a result of which the machine learning algorithm generates a predictive indication of the interaction of a given user with a given digital element from the plurality of digital elements;
11. - the use of a loss function that can penalize a predicted user interaction if that predicted interaction differs from the corresponding label, thereby training the machine learning algorithm to find, for a given user, previously undisplayed digital elements in a set of digital elements that are associated with a previously unknown source that the user is likely to interact with.
12. 2. The method according to claim 1, wherein finding the training set of past digital elements involves finding the training set of past digital elements in a predetermined period.
13. 3. The method according to claim 2, wherein finding the training set of past digital elements in a predetermined period involves finding the latest past digital elements in a predetermined period.
14. 4. The method according to paragraph 2, in which the steps of finding, forming, transmitting and applying are performed at the first iteration of training, and the method at the second iteration of training following the first iteration of training additionally provides:
15. - finding in the user log associated with a given user a second training set of past digital elements from a plurality of digital elements of the digital recommendation platform in a predetermined period, wherein another given past digital element from the second training set of past digital elements is associated with an indication of the interaction of the given user with the given past digital element in the predetermined period;
16. - forming a second training data set containing a second set of training digital objects, wherein forming another specified training digital object from the second set of training digital objects includes:
17. - scanning a second training set of past digital items arranged in chronological order to find, for another given past time in a predetermined period, another given target past digital item from the target source, which is the first digital item from among the digital items associated with the target source with which the given user interacted before the given past time in the predetermined period;
18. - forming another training subset of past digital items containing past digital items with which a given user interacted before interacting with another given target past digital item in a predetermined period;
19. - forming another given training digital object containing a vector of a given user, a combination of vectors of elements from another training subset of past digital elements associated with another given target past digital element, and a label containing an indication of the interaction of a given user with another given target past digital element associated with a target source;
20. - transmitting each of the second set of training digital objects to the machine learning algorithm, as a result of which the machine learning algorithm generates another predictive indication of the interaction of a given user with a given digital element from the set of digital elements;