Search engine using joint learning for multi-label classification

Patent metadata

Field	Value
Publication number	US-2023252095-A1
Application number	US-202318134412-A
Country	US
Kind code	A1
Filing date	Apr 13, 2023
Priority date	Oct 15, 2019
Publication date	Aug 10, 2023
Grant date	—

Abstract

A method of configuring a search engine to classify a search query includes receiving a search query data set, the search query data set comprising a plurality of search queries, defining a first set of candidate labels and a second set of candidate labels according to the search queries in the search query data set, concatenating the first set of candidate labels with the second set of candidate labels to generate a concatenated candidate label set, generating a compatibility matrix comprising a similarity between the concatenated candidate label set and the search query data set, and training a classification network according to the compatibility matrix.

First claim

1 - 20 . (canceled)

21 . A method of configuring a search engine to classify a search query, the method comprising:
defining a matrix of first candidate labels based on a plurality of categories associated with user navigation histories; defining a matrix of second candidate labels based on a plurality of user intents in the user navigation histories; concatenating the matrix of the first candidate labels with the matrix of the second candidate labels to generate a concatenated candidate label vector; determining a respective similarity between the concatenated candidate label vector and each of a plurality of search queries; and training a classification network according to the determined similarities.

22 . The method of claim 3 , further comprising:
deriving a plurality of search queries from the user navigation histories; converting each of the plurality of search queries into respective embeddings; and generating a compatibility matrix indicative of a similarity between the concatenated candidate label vector and the embeddings.

23 . The method of claim 3 , wherein training the classification network according to the determined similarities comprises training a first neural network for determining a category of a new search query and a second neural network for determining a user intent of the new search query.

24 . The method of claim 3 , further comprising:
determining a plurality of items included in the user navigation histories; determining a plurality of categories associated with the plurality of items; and identifying a plurality of user intents in the user navigation histories.

25 . The method of claim 4 , wherein the first candidate labels comprise category labels from the plurality of categories associated with the plurality of items.

26 . The method of claim 4 , wherein the second candidate labels comprise user intent statements from the plurality of user intents.

27 . The method of claim 3 , wherein each determined similarity comprises a cosine similarity between the concatenated candidate label vector and one of the plurality of search queries.

28 . A system for configuring a search engine to classify a search query, the system comprising:
a non-transitory, computer-readable memory storing instructions; and a processor configured to execute the instructions to:
define a matrix of first candidate labels based on a plurality of categories associated with user navigation histories;
define a matrix of second candidate labels based on a plurality of user intents in the user navigation histories;
concatenate the matrix of the first candidate labels with the matrix of the second candidate labels to generate a concatenated candidate label vector;
determine a respective similarity between the concatenated candidate label vector and each of a plurality of search queries; and
train a classification network according to the determined similarities.

29 . The system of claim 8 , wherein the processor is further configured to:
deriving a plurality of search queries from the user navigation histories; converting each of the plurality of search queries into respective embeddings; and generating a compatibility matrix comprises a similarity between the concatenated candidate label vector and the embeddings.

30 . The system of claim 8 , wherein training the classification network according to the determined similarities comprises training a first neural network for determining a category of a new search query and a second neural network for determining a user intent of the new search query.

31 . The system of claim 8 , wherein the processor is further configured to:
determining a plurality of items included in the user navigation histories; determining a plurality of categories associated with the plurality of items; and identifying a plurality of user intents in the user navigation histories.

32 . The system of claim 31 , wherein the first candidate labels comprise category labels from the plurality of categories associated with the plurality of items.

33 . The system of claim 31 , wherein the second candidate labels comprise user intent statements from the plurality of user intents.

34 . The system of claim 8 , wherein each determined similarity comprises a cosine similarity between the concatenated candidate label vector and one of the plurality of search queries.

35 . A method for responding to a user search request, the method comprising:
defining a matrix of first candidate labels based on a plurality of categories associated with user navigation histories; defining a matrix of second candidate labels based on a plurality of user intents in the user navigation histories; concatenating the matrix of the first candidate labels with the matrix of the second candidate labels to generate a concatenated candidate label vector; determining a respective similarity between the concatenated candidate label vector and each of a plurality of search queries; training a classification network according to the determined similarities; receiving, by a server, a user search query; applying, by the server, the trained classification network to the user search query to identify at least one of a user intent or an item category; and providing, by the server, a response to the user search query according to the at least one of a user intent or an item category.

36 . The method of claim 35 , further comprising:
deriving a plurality of search queries from the user navigation histories; converting each of the plurality of search queries into respective embeddings; and generating a compatibility matrix comprises a similarity between the concatenated candidate label vector and the embeddings.

37 . The method of claim 35 , wherein training the classification network according to the determined similarities comprises training a first neural network for determining a category of a new search query and a second neural network for determining a user intent of the new search query.

38 . The method of claim 35 , further comprising:
determining a plurality of items included in the user navigation histories; determining a plurality of categories associated with the plurality of items; and identifying a plurality of user intents in the user navigation histories.

39 . The method of claim 38 , wherein the first candidate labels comprise category labels from the plurality of categories associated with the plurality of items.

40 . The method of claim 38 , wherein the second candidate labels comprise user intent statements from the plurality of user intents.

Assignees

• Home Depot Product Authority Llc

Inventors

• Ahmadvand Ali
• Kallumadi Surya
• Javed Faizan

Classifications

• G06N3/091

  Active learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06F16/9535 Primary

  Search customisation based on user profiles and personalisation · CPC title

• G06F18/2148

  characterised by the process organisation or structure, e.g. boosting cascade · CPC title