Concept refinement using concept activation vectors

Patent metadata

Field	Value
Publication number	US-12154160-B2
Application number	US-202218050378-A
Country	US
Kind code	B2
Filing date	Oct 27, 2022
Priority date	Oct 27, 2022
Publication date	Nov 26, 2024
Grant date	Nov 26, 2024

Abstract

A search query and a concept comprising a feature are received from a user. Positive and negative item sets are identified, where items of the positive item set comprise the feature and items of the negative item set do not comprise the feature. A hyperplane located between the positive and negative item sets is generated using a machine learning model. A concept activation vector (CAV) orthogonal to the hyperplane is determined. The CAV is used to produce a modified search query vector. Based on comparing the modified search query vector with item vectors, a second set of search results is provided.

First claim

We claim:

1. Computer storage media storing computer-executable instructions that, when executed by at least one processor, cause the at least one processor to perform operations comprising:
accessing a first set of search results for a search query received at a search engine, the first set of search results comprising a first set of items from an item index comprising items having corresponding item vectors, the first set of items identified based on comparing a search query vector with the item vectors;
receiving a concept input comprising a feature;
identifying, from the item index, a positive item set and a negative item set, wherein the positive item set comprises a first item that includes the feature and the negative item set comprises a second item that does not include the feature;
generating, using a machine learning model, a concept activation vector based on the positive item set and the negative item set;
modifying the search query vector using the concept activation vector to generate a modified search query vector; and
providing a second set of search results comprising a second set of items identified from the item index based on comparing the modified search query vector with the item vectors.

2. The computer-implemented method of claim 1 , wherein the concept input is a text string corresponding to the feature, and wherein the method further comprises:
identifying the first item by determining that the first item comprises the text string in a description of the first item; and
identifying the second item by determining that the second item does not comprise the text string in a description of the second item.

3. The computer-implemented method of claim 1 , wherein the concept input is a text string corresponding to the feature, and wherein the method further comprises:
identifying the first item by determining that a title of the first item comprises the text string; and
identifying the second item by determining that a title of the second item does not comprise the text string.

4. The computer-implemented method of claim 1 , wherein the concept input is an image comprising the feature, and wherein the method further comprises:
identifying the first item by determining that a description of the first item comprises the feature; and
identifying the second item by determining that a description of the second item does not comprise the feature.

5. The computer-implemented method of claim 4 , further comprising determining a text string corresponding to the feature in the image, wherein the description of the first item comprises the text string, and wherein the description of the second item does not comprise the text string.

6. The computer-implemented method of claim 1 , wherein the first item is identified based on a positive item set threshold, and wherein the second item is identified based on a negative item set threshold.

7. The computer-implemented method of claim 1 , wherein generating the concept activation vector using the machine learning model comprises:
determining a hyperplane located between item vectors of the positive item set and item vectors of the negative item set; and
determining a direction of the concept activation vector, wherein the direction of the concept activation vector is orthogonal to the hyperplane and toward the item vectors of the positive item set.

8. The computer-implemented method of claim 7 , wherein generating the concept activation vector using the machine learning model further comprises:
determining a length of the concept activation vector based on an amount of separation between the item vectors of the positive item set and the item vectors of the negative item set.

9. A system comprising:
at least one processor; and
computer storage media storing computer-executable instructions that, when executed by the at least one processor, cause the at least one processor to perform operations comprising:
providing a search query;
receiving a first set of search results in response to the search query, the first set of search results comprising a first set of items from an item index comprising items having corresponding item vectors, the first set of items identified based on comparing a search query vector with the item vectors;
providing a concept input comprising a feature, wherein the providing the concept input causes generation of a concept activation vector determined by:
identifying, from the item index, a positive item set and a negative item set, wherein the positive item set comprises a first item that includes the feature and the negative item set comprises a second item that does not include the feature; and
generating, using a machine learning model, the concept activation vector based on the positive item set and the negative item set; and

receiving a second set of search results based on a modified search query vector, wherein the modified search query vector is determined by modifying the search query vector using the concept activation vector.

10. The system of claim 9 , wherein the concept input is a text string corresponding to the feature, wherein the first item is identified by determining that the first item comprises the text string in a description of the first item, and wherein the second item is identified by determining that the second item does not comprise the text string in a description of the second item.

11. The system of claim 10 , wherein the concept input is a text string corresponding to the feature, wherein the first item is identified by determining that a title of the first item comprises the text string, and wherein the second item is identified by determining that a title of the second item does not comprise the text string.

12. The system of claim 9 , wherein the concept input is an image comprising the feature, wherein the first item is identified by determining that a first image associated with the first item comprises the feature, and wherein the second item is identified by determining that a second image associated with the second item does not comprise the feature.

13. The system of claim 9 , wherein the machine learning model comprises a support-vector machine.

14. The system of claim 9 , wherein the concept activation vector is further generated by:
determining a hyperplane located between item vectors of the positive item set and item vectors of the negative item set; and
determining a direction of the concept activation vector, wherein the direction of the concept activation vector is orthogonal to the hyperplane and toward the item vectors of the positive item set.

15. The system of claim 14 , wherein the concept activation vector is further generated by:
determining a length of the concept activation vector based on an amount of separation between the item vectors of the positive item set and the item vectors of the negative item set.

16. A computer-implemented method comprising:
accessing a positive item set and a negative item set, wherein the positive item set and the negative item set are determined by:
accessing a first set of search results for a search query received at a search engine, the first set of search results comprising a first set of items from an item index comprising items having corresponding item vectors, the first set of items identified based on comparing a search query vector with the item vectors;
receiving a concept input comprising a feature; and
identifying, from the item index, the positive item set and the negative item set, wherein the positive item set comprises a first item that includes the feature and the negative i

Assignees

• Ebay Inc

Inventors

• Basin Dmitry
• Ben-Shaul Ido

Classifications

• G06Q30/0627

  by specifying product or service characteristics, e.g. product dimensions · CPC title

• G06N20/10

  using kernel methods, e.g. support vector machines [SVM] · CPC title

• G06Q30/0631 Primary

  Recommending goods or services · CPC title

• G06F16/9532 Primary

  Query formulation · CPC title

• G06F16/9535 Primary

  Search customisation based on user profiles and personalisation · CPC title