SYSTEM AND METHOD OF FORMING TRAINING SET FOR MACHINE LEARNING ALGORITHM Russian patent published in 2021 - IPC G06N20/00 

FIELD: machine learning.

SUBSTANCE: invention relates to methods and a system for forming a set of training objects for training a machine learning algorithm (hereinafter – MLA) performed by a server. The method is to perform obtaining of the first query sent to the server and associated with the first set of search results associated with the user action parameter. Then, based on the terms of the first query, a set of queries previously sent to the server that differ from the first query by the specified number of terms is obtained. Then, for a set of queries, a set of search results is got that differ from the search results in the first set of search results. After that, the similarity score of the first query and queries from the query set is calculated based on the first set of search results, other search results sets, and user action parameters in the first and other search results sets. Then, subsets of queries from a set of queries are determined based on a similarity score less than a predefined similarity threshold. Finally, a set of training objects is formed to be used as negative training examples for MLA training, with each training object containing the first query, another query from a subset of queries, and a similarity score of the first query and the other query.

EFFECT: technical result is increased efficiency of MLA training.

27 cl, 5 dwg