System and a method of training a machine-learning models for search results ranking
A method and system for training a machine-learning algorithm (MLA), including first and second machine-learning (ML) models to rank digital documents at a digital platform. The method comprises: receiving first historical data of past searches conducted by the given user over a given past period; receiving second historical data, a given portion of the second historical data including data of past searches conducted by the given user over a given past user session; jointly training both the first and second ML models to rank the in-use digital documents, the training comprising: training, based on the first historical data, the first ML model to generate a vector representation of the first historical data; and training, based on the vector representation of the first historical data and the second historical data, the second ML model to determine a respective likelihood value of the given user interacting with a given in-use digital document.
1. A computer-implementable method of training a machine-learning algorithm (MLA) to rank in-use digital documents at a digital platform, the MLA comprising a first ML model and a second ML model, the method being executable by a processor, the method comprising:
receiving, by the processor, a first historical data including: (i) a first plurality of training queries submitted by a given user to the digital platform over a given past period; and (ii) respective sets of training digital documents generated, by the digital platform, during the given past period in response to submitting thereto each one of the first plurality of training queries;
receiving, by the processor, a second historical data, a given portion of the second historical data including: (i) a second plurality of training queries submitted by the given user to the digital platform over a given past training user session, the given past training user session being shorter and more recent than the given past period; and (ii) respective sets of training digital documents generated, by the digital platform, during the given past training user session in response to submitting thereto each one of the second plurality of training queries,
a given training digital document generated during any one of the given past period and the given past training user session including a respective indication of a user interaction of the given user with the given training digital document; and
jointly training, by the processor, both the first and second ML models to rank the in-use digital documents, the training comprising:
organizing, by the processor, the first historical data into a first plurality of training digital objects, a given one of which includes: (i) a given one of the first plurality of training queries; and (ii) the respective set of training digital documents generated in response to the given one of the first plurality of training queries during the given past period;
feeding, by the processor, the first plurality of training digital objects to the first ML model to train the first ML model to generate a vector representation of the first historical data;
generating, by the processor, based on the second historical data, a second plurality of training digital objects, a given one of which includes: (i) a given one of the second plurality of training queries; (ii) the respective set of training digital documents generated in response to the given one of the second plurality of training queries during the given training user session; and (iii) the vector representation of the first historical data; and
feeding, by the processor, the second plurality of training digital objects to the second ML model to train the second ML model to determine a respective likelihood value of the given user interacting with a given in-use digital document.
2. The method of claim 1 , wherein a number of members of the first plurality of training digital objects is greater than a number of members of the second plurality of training digital objects.
3. The method of claim 1 , wherein the user interaction of the given user with the given training digital document comprises at least one: (i) selecting the given training digital document from a respective set of training digital documents; (ii) dwelling on the given digital document; (iii) adding the given training digital document in favorites; (iv) engaging with the given training digital document for longer than a threshold engagement period; and (v) saving at least a portion of content of the given training digital document.
4. The method of claim 1 , further comprising:
receiving, by the processor, a given in-use query submitted by the given user to the digital platform during a current in-use user session;
retrieving, by the processor, a respective set of in-use digital documents responsive to the given in-use query;
retrieving, by the processor, in-use historical data including: (i) a plurality of past in-use queries submitted by the given user to the digital platform over the current in-use user session prior to submitting the given in-use query; (ii) respective sets of past in-use training digital documents generated, by the digital platform, in response to submitting thereto each one of the plurality of past in-use training queries, a given past in-use training digital document of the respective sets of past in-use digital documents including the respective indication of the user interaction of the given user with the given past in-use training digital document;
generating, by the processor, an in-use digital object including: (i) the given in-use query; (ii) the respective set of in-use digital documents; (iii) the plurality of past in-use queries; (iv) respective sets of past in-use training digital documents; and (v) the vector representation of the first historical data;
feeding, by the processor, the in-use digital object to the second ML model of the MLA to determine, for each one of the set of in-use digital documents, the respective likelihood value of the given user interacting therewith; and
ranking, by the processor, each one of the set of in-use digital documents in accordance with the respective likelihood values associated therewith.
5. The method of claim 4 , further comprising, prior to the retrieving the vector representation of the first historical data, updating, by the processor, the first historical data.
6. The method of claim 5 , wherein the updating comprises at least one of shifting and extending the given past period towards a moment of submitting the given in-use query.
7. The method of claim 5 , wherein the updating is executed at a predetermined frequency.
8. The method of claim 1 , wherein each one of the first and second ML models is a neural network having an encoder-decoder architecture.
9. The method of claim 8 , wherein the neural network is Transformer-based neural network.
10. A server for training a machine-learning algorithm (MLA) to rank in-use digital documents at a digital platform, the MLA comprising a first ML model and a second ML model, the server comprising a non-transitory computer-readable memory storing instructions; and a processor, which, upon executing the instructions, is configured to:
receive a first historical data including: (i) a first plurality of training queries submitted by a given user to the digital platform over a given past period; and (ii) respective sets of training digital documents generated, by the digital platform, during the given past period in response to submitting thereto each one of the first plurality of training queries;
receive a second historical data, a given portion of the second historical data including: (i) a second plurality of training queries submitted by the given user to the digital platform over a given past training user session, the given past training user session being shorter and more recent than the given past period; and (ii) respective sets of training digital documents generated, by the digital platform, during the given past training user session in response to submitting thereto each one of the second plurality of training queries,
a given training digital document generated during any one of the given past period and the given past training user session including a respective indication of a user interaction of the given user with the given training digital document; and
jointly train both the first and second ML models to rank the in-use digital documents, by:
organizing the first historical data into a first plurality of training digital objects, a given one of which includes: (i) a given one of the first plurality of training queries; and (ii) the respective set of training digital documents generated in response to the given one of the first plurality of training queries during the given past period;
feeding the first plurality of training digital objects to the first ML model to train the first ML model to generate a vector representation of the first historical data;
generating, based on the second historical data, a second plurality of training digital objects, a given one of which includes: (i) a given one of the second plurality of training queries; (ii) the respective set of training digital documents generated in response to the given one of the second plurality of training queries during the given training user session; and (iii) the vector representation of the first historical data; and
feeding the second plurality of training digital objects to the second ML model to train the second ML model to determine a respective likelihood value of the given user interacting with a given in-use digital document.
11. The server of claim 10 , wherein a number of members of the first plurality of training digital objects is greater than a number of members of the second plurality of training digital objects.
12. The server of claim 10 , wherein the user interaction of the given user with the given training digital document comprises at least one: (i) selecting the given training digital document from a respective set of training digital documents; (ii) dwelling on the given digital document; (iii) adding the given training digital document in favorites; (iv) engaging with the given training digital document for longer than a threshold engagement period; and (v) saving at least a portion of content of the given training digital document.
13. The server of claim 10 , wherein the processor is further configured to:
receive a given in-use query submitted by the given user to the digital platform during a current in-use user session;
retrieve a respective set of in-use digital documents responsive to the given in-use query;
retrieve in-use historical data including: (i) a plurality of past in-use queries submitted by the given user to the digital platform over the current in-use user session prior to submitting the given in-use query; (ii) respective sets of past in-use training digital documents generated, by the digital platform, in response to submitting thereto each one of the plurality of past in-use training queries, a given past in-use training digital document of the respective sets of past in-use digital documents including the respective indication of the user interaction of the given user with the given past in-use training digital document;
generate an in-use digital object including: (i) the given in-use query; (ii) the respective set of in-use digital documents; (iii) the plurality of past in-use queries; (iv) respective sets of past in-use training digital documents; and (v) the vector representation of the first historical data;
feed the in-use digital object to the second ML model of the MLA to determine, for each one of the set of in-use digital documents, the respective likelihood value of the given user interacting therewith; and
rank each one of the set of in-use digital documents in accordance with the respective likelihood values associated therewith.
14. The server of claim 13 , wherein, prior to retrieving the vector representation of the first historical data, the processor is further configured to update the first historical data.
15. The server of claim 14 , wherein the processor is configured to update the first historical data by executing at least one of shifting and extending the given past period towards a moment of submitting the given in-use query.
16. The server of claim 15 , wherein the processor is configured to update the first historical data at a predetermined frequency.
17. The server of claim 10 , wherein each one of the first and second ML models is a neural network having an encoder-decoder architecture.
18. The server of claim 17 , wherein the neural network is Transformer-based neural network.