US 12,306,886 B2
Method and a server for generating a machine learning model
Mikhail Dektyarev, Vladimir (RU); and Alexander Safronov, Moscow (RU)
Assigned to Y.E. Hub Armenia LLC, Yerevan (AM)
Filed by Yandex Music LLC, Moscow (RU)
Filed on Jul. 5, 2023, as Appl. No. 18/218,395.
Claims priority of application No. 2022118288 (RU), filed on Jul. 5, 2022.
Prior Publication US 2024/0012861 A1, Jan. 11, 2024
Int. Cl. G06F 16/9535 (2019.01)
CPC G06F 16/9535 (2019.01) 10 Claims
OG exemplary drawing
 
1. A method of generating a machine learning model, the machine learning model to be used for generating a digital item recommendation to a user of a recommendation system, the user being one of a plurality of users, the recommendation system being executed by a server, the user being associated with an electronic device, the electronic device being communicatively connectable with the server over a communication network, the method executable by the server, the method comprising:
receiving user-item interaction data indicative of previous interactions between the plurality of users and a plurality of digital items available for recommendation to the plurality of users within the recommendation system;
generating based on the user-item interaction data, the machine learning model for predicting scores indicative of a likelihood that a given user of the plurality of users will interact with a given digital item of the plurality of digital items,
the machine learning model including a plurality of item-specific decision-tree (ISDT) sub-models, a given ISDT sub-model of the plurality of ISDT sub-models being associated with the given digital item for generating a score indicative of the likelihood that the given user will interact with the given digital item;
the generating including generating for the given ISDT sub-model a training set including:
a training target set for the given digital item containing the user-item interaction data associated with the given digital item;
a training input set for the given digital item containing the user-item interaction data associated with a subset of digital items of the plurality of digital items, the subset of digital items excluding the given digital item,
the subset of digital items to be used as training features for generating the given ISDT sub-model such that previous user-item interactions between the plurality of users and the subset of digital items are used as values for respective training features;
the generating including generating, a given plurality of decision trees (DTs) using the training set, the given plurality of DTs forming the given ISDT sub-model,
a given one from the given plurality of DTs having feature nodes and leaf nodes, each of the feature nodes corresponding to a digital item of the subset of digital items, and each of the leaf nodes comprising a score indicative of the likelihood that the given user will interact with the given digital item.