Method of and server for generating meta-feature for ranking documents

1. A computer-implemented method for generating a meta-feature for ranking documents by a machine learning algorithm (MLA) executed by a server, the method executable by the server, the method comprising:

acquiring, by the server, an indication of a past query having been submitted on the server;

acquiring, by the server, a set of past documents, the set of past documents having been presented as search results in response to the past query, each past document of the set of past documents including:

a first plurality of features having been used for ranking past documents in the set of past documents, and

respective values for the first plurality of features;

generating, by the server, the meta-feature, a respective value of the meta-feature for a given past document being a relative value between a respective value of a given feature of the first plurality of features for the given past document and at least one other respective value of the given feature associated with other ones of the set of past documents;

validating, by the server, the meta-feature based on a usefulness thereof for ranking future search engine results pages (SERPs),

the usefulness of the meta-feature being determined by comparing, by the processor, user interactions with: (i) a current ranking of the set of past documents and (ii) a new ranking of the set of past documents, the new ranking being based at least on the meta-feature;

in response to the usefulness of the meta-feature being above a predetermined threshold, generating a plurality of training objects for training the MLA for ranking future SERPs,

a given training object of the plurality of training objects including: (i) the indication of the past query; (ii) the given past document of the set of past documents; (iii) the respective values of the first plurality of features associated therewith; and (iv) the respective value of the meta-feature associated therewith.

2. The method of claim 1 , wherein

the set of past documents is associated with past user interactions; and wherein the method further comprises prior to the validating:

determining, by the server, the threshold based on the past user interactions with the set of past documents; and wherein the validating comprises:

receiving, from a plurality of electronic device connected to the server, a current query, the current query being similar to the past query;

generating, by the server, a respective set of current documents relevant to the current query, each current document of the respective set of current documents including the first plurality of features and the meta-feature, the respective set of current documents being at least a subset of the set of past documents;

ranking, by the MLA, the respective set of current documents based at least in part on the first plurality of features and the meta-feature to obtain a respective final ranked list of documents;

transmitting, by the server to the plurality of electronic devices, a respective SERP including the respective final ranked list of documents;

receiving, by the server from the plurality of electronic devices, at least one respective user interaction with the respective SERP; and

determining, by the server, the usefulness of the meta-feature based on:

the respective user interactions with the respective SERPs.

3. The method of claim 1 , wherein the respective value of the meta-feature for the given past document is based on the respective value of a given feature of the first plurality of features for the given past document and a value of a parameter associated with the set of past documents, the value of the parameter including at least one of:

a respective preliminary rank of the given past document, and

another value of another feature associated with one of the given past document and at least one other document.

4. The method of claim 2 , wherein the method further comprises, prior to the generating the meta-feature:

ranking, by a second MLA executed by the server, the set of past documents based on at least a first feature of the first plurality of features to obtain a preliminary ranked list of documents, each past document of the set of past documents having the respective preliminary rank.

5. The method of claim 4 , wherein the generating the meta-feature is further based on:

a respective value of another given feature of the first plurality of features for the given past document, and

a value of a second parameter associated with the set of past documents.

6. The method of claim 5 , wherein the value of the second parameter is at least one of:

the respective preliminary rank of the given past document,

a value of the other given feature of at least one other document of the set of past documents, and

another value of another feature associated with one of the given past document and at least one other document.

7. The method of claim 3 , wherein the value of the parameter for the given past document in the set of past documents is an average value of the given feature for the set of past documents.

8. The method of claim 3 , further comprising: repeating the method for a set of past queries, a respective set of past documents having been provided as respective search results in response to a respective past query of the set of past queries, the respective set of past documents being associated with respective past user interactions.

9. The method of claim 8 , wherein the determining the threshold comprises:

applying, by the server, a user engagement metric on the respective sets of past documents based on the respective past user interactions to obtain the threshold; and wherein

the determining the usefulness of the meta-feature comprises:

applying, by the server, a current user engagement metric on the respective SERPs based on the respective user interactions with the respective SERPs to obtain the usefulness.

10. The method of claim 7 , wherein the meta-feature is a first meta-feature of a set of meta-features; and wherein

the generating the first meta-feature further comprises generating each respective meta-feature of the set of meta-features, each respective value of the respective meta-feature being generated based on:

a respective value of a respective feature of the first plurality of features, and

a respective parameter associated with the respective set of past documents; and

the determining to use the first-meta feature is executed further in response to the current user engagement metric of the first meta-feature being above respective current user engagement metrics of remaining meta-features of the set of meta-features.

11. The method of claim 2 , wherein the given feature is one of:

a query-dependent feature, and

a query-independent feature.

12. The method of claim 11 , wherein the respective value of the query-independent feature for the given past document is one of:

past values for the query-independent feature, and

predicted values for the query independent feature.

13. A computer-implemented method for generating a meta-feature for ranking documents by a machine learning algorithm (MLA) executed by a server, the MLA having been trained to generate the meta-feature for ranking the documents in response to a given query, the method executable by the server, the method comprising:

receiving, from an electronic device connected to the server, a new query, the MLA not having been trained to rank documents based at least in part on the meta-feature for the new query;

generating, by the server, a set of current documents relevant to the new query, each current document of the respective set of current documents including a first plurality of features;

generating, by the MLA, the meta-feature, a respective value of the meta-feature for a given current document being a relative value between a respective predicted value of a given feature of the first plurality of features for the given current document and at least one other respective predicted value of the given feature associated with other ones of the set of current documents;

ranking, by the MLA, the set of current documents based at least in part on the first plurality of features and the meta-feature to obtain a respective final ranked list of documents; and

transmitting, by the server to the electronic device, a respective SERP including the respective final ranked list of documents.

14. The method of claim 13 , further comprising:

during a training phase:

acquiring, by the server, a set of past queries, each query of the set of past queries having been previously submitted on the server;

acquiring, for each query of the set of past queries, a respective set of past documents, the respective set of past documents having been presented as respective search results in response to the respective query, each past document of the respective set of past documents having:

a first plurality of features, and

respective values for the first plurality of features;

generating, for each set of past documents, the meta-feature, a respective value of the meta-feature for a given past document of the respective set of past documents being a relative value between a respective value of a given feature of the first plurality of features for the given past document, and at least one other value of the given feature for other ones in the respective set of past documents;

validating, by the server, the meta-feature based on a usefulness thereof for ranking future search engine results pages (SERPs) in response to current queries, each of the current query being one of the set of past queries,

the usefulness of the meta-feature being determined by comparing, by the processor, user interactions with: (i) a current ranking of the respective set of past documents and (ii) a new ranking of the respective set of past documents, the new ranking being based at least on the meta-feature;

in response to the usefulness of the meta-feature being above a predetermined threshold:

training, by the server, the MLA to generate the meta-feature.

15. The method of claim 14 , wherein the validating comprises:

receiving, from a plurality of electronic device connected to the server, the current queries;

generating, by the server, a respective set of current documents relevant to each one of the current queries, each current document of the respective set of current documents including the first plurality of features and the meta-feature;

ranking, by the MLA, the respective sets of current documents based at least in part on the first plurality of features and the meta-feature to obtain the respective final ranked list of documents;

transmitting, by the server to the plurality of electronic devices, respective SERPs, each respective SERPs including the respective final ranked list of documents;

receiving, by the server from the plurality of electronic devices, at least one respective user interaction with the respective SERP; and

determining, by the server, the usefulness of the meta-feature based on:

the respective user interactions with the respective SERPs.

16. The method of claim 15 , wherein the respective set of past documents is associated with respective past user interactions; and wherein

the method further comprises, prior to the validating the meta-feature:

applying, by the server, a user engagement metric on the respective sets of past documents based on the respective past-user interactions to obtain the threshold; and wherein

the determining the usefulness of the meta-feature comprises:

applying, by the server, a current user engagement metric on the respective SERPs based on the respective user interactions with the respective SERPs to obtain the usefulness.

17. The method of claim 16 , wherein the MLA is trained to generate the meta-feature based on:

the meta-feature,

the respective SERP, the respective SERP including the respective final ranked list of documents having been generated based in part on the meta-feature, and

the respective user interactions with the respective SERP.

18. A computer-implemented method for ranking documents in response to a given query using a meta-feature by a machine learning algorithm (MLA) executed by a server, the method executable by the server, the method comprising:

receiving, from an electronic device connected to the server, a given query;

generating, by the server, a set of current documents relevant to the new query, each current document of the respective set of current documents including a first plurality of features;

ranking, by the MLA, the set of current documents based on at least a portion of the first plurality of features to obtain a preliminary ranked list of documents;

generating, by the MLA, the meta-feature, a respective value of the meta-feature for a given current document in the preliminary ranked list of documents being a relative value between a respective value of a given feature of the first plurality of features for the given current document and a respective preliminary ranking score thereof in the preliminary ranked list of documents;

ranking, by the MLA, the preliminary ranked list of documents based on at least the meta-feature to obtain a respective final ranked list of documents; and

transmitting, by the server to the electronic device, a respective SERP including the respective final ranked list of documents in response to the given query.