MULTISTAGE TRAINING OF MACHINE LEARNING MODELS FOR RANKING RESULTS Russian patent published in 2024 - IPC G06F16/00 

FIELD: physics.

SUBSTANCE: present technology relates to machine learning methods, in particular to methods and systems for training and using transformer-based machine learning models for ranking search results. Method includes training an MLA algorithm at a first step to determine past interactions of a user with past digital documents based on a first set of training objects containing past digital documents, determined by the interactive search platform in response to sending it by the user corresponding past requests. Method also includes training the MLA algorithm at the second stage to determine values of the probability of user interaction with digital documents of the use stage based on the second set of training objects, containing only past digital documents with which the user interacted, and past requests associated with them.

EFFECT: technical result is to provide more effective training of the training model using a limited amount of training data based on the search results ranking transformer, as well as improved quality of prediction of relevance of search result for a specific user, that is, improved personalized ranking.

16 cl, 6 dwg