Identifying entities using a deep-learning model

Abstract

In one embodiment, a method includes retrieving a first vector representation of a first entity, with which a user has interacted, and a second vector representation of a second entity, with which the user has not interacted. The first and second vector representations are determined using an initial deep-learning model. A first similarity score is computed between a vector representation of the user and the first vector representation, and a second similarity score is computed between the vector representation of the user and the second vector representation. The second vector representation is updated if the second similarity score is greater than the first similarity score using the initial deep-learning model. An updated deep-learning model is generated based on the initial deep-learning model and on the updated second vector representation.

What is claimed is: 
     
         1 . A method comprising, by one or more computing devices:
 retrieving a first vector representation of a first entity that a user has interacted with, wherein the first vector representation is generated using an initial deep-learning model;   retrieving a second vector representation of a second entity that the user has not interacted with, wherein the second vector representation is generated using the initial deep-learning model;   determining a first similarity score between the first vector representation and a vector representation of the user;   determining a second similarity score between the second vector representation and the vector representation of the user;   upon determining that the second similarity score is greater than the first similarity score, updating, using the initial deep-learning model, the second vector representation; and   generating an updated deep-learning model based on the initial deep-learning model and the updated second vector representation.   
     
     
         2 . The method of  claim 1 , wherein the retrieving of the first vector representation comprises:
 accessing a set of vector representations of entities that the user has interacted with, wherein the set comprises the first vector representation and a plurality of additional vector representations, each additional vector representation corresponding to a respective entity that the user has interacted with;   selecting, from the set, the first vector representation; and   removing the first vector representation from the set.   
     
     
         3 . The method of  claim 2 , wherein the first vector representation is selected randomly from the set of vector representations. 
     
     
         4 . The method of  claim 1 , wherein retrieving the second vector representation comprises:
 accessing a set comprising a plurality of additional vector representations, each additional vector representation corresponding to a respective entity that the user has not interacted with, wherein the set further comprises the second vector representation; and   selecting, from the set, the second vector representation and one or more additional vector representations.   
     
     
         5 . The method of  claim 4 , wherein a number of the selected one or more additional vector representations is based on a loss function of the initial deep-learning model. 
     
     
         6 . The method of  claim 4 , further comprising:
 determining, for each additional vector representation, a respective similarity score between the additional vector representation and the vector representation of the user;   assigning a respective ranking to the first vector representation and to each additional vector representation based on the first similarity score and the respective similarity scores; and   updating one or more of the additional vector representations based on the respective rankings.   
     
     
         7 . The method of  claim 1 , further comprising:
 retrieving a plurality of additional vector representations, each additional vector representation corresponding to a respective additional entity that the user has interacted with; and   combining the plurality of additional vector representations to generate the vector representation of the user.   
     
     
         8 . The method of  claim 1 , wherein updating the second vector representation comprises iteratively updating the second vector representation until the second similarity score is less than the first similarity score. 
     
     
         9 . The method of  claim 1 , further comprising:
 retrieving a third vector representation corresponding to a third entity that that the user has not interacted with; and   predicting a relevance to the user of the third entity using the updated deep-learning model.   
     
     
         10 . The method of  claim 9 , wherein the third vector representation is generated by the updated deep-learning model. 
     
     
         11 . The method of  claim 1 , further comprising accessing a multi-dimensional embedding space, wherein the embedding space comprises:
 a user embedding corresponding to the vector representation of the user;   a first embedding corresponding to the first vector representation;   a second embedding corresponding to the second vector representation; and   a plurality of additional embeddings, each additional embedding corresponding to a respective additional vector representation of an additional entity.   
     
     
         12 . The method of  claim 11 , further comprising identifying a third entity that the user has not interacted with as relevant to the user by applying a search algorithm to the embedding space, wherein the search algorithm determines that an embedding corresponding to a third vector representation of the third entity is within a threshold distance of the user embedding in the embedding space. 
     
     
         13 . The method of  claim 12 , further comprising:
 sending, to a client system of the user, instructions for presenting the identified third entity to the user.   
     
     
         14 . The method of  claim 1 , wherein the user is a user of a social-networking system, and wherein the method further comprises:
 accessing a social graph of the social-networking system, the social graph comprising a plurality of nodes and a plurality of edges connecting the nodes, each of the edges between two of the nodes representing a single degree of separation between them, the nodes comprising:
 a first node corresponding to the user; and 
 a plurality of second nodes, each second node corresponding to a respective entity. 
   
     
     
         15 . The method of  claim 14 , wherein the user has interacted with the first entity via a social-networking action of the user, wherein the social-networking action is taken with respect to the first node and a second node corresponding to the first entity. 
     
     
         16 . The method of  claim 15 , wherein the social-networking action represents an expression of affinity for the first entity. 
     
     
         17 . The method of  claim 1 , wherein either one or both of the first entity or the second entity comprise a page hosted by a social-networking system. 
     
     
         18 . The method of  claim 1 , wherein the first vector representation, the second vector representation, and the vector representation of the user each comprise a respective d-dimensional intensity vector. 
     
     
         19 . One or more computer-readable non-transitory storage media embodying software that is operable when executed to:
 retrieve a first vector representation of a first entity that a user has interacted with, wherein the first vector representation is generated using an initial deep-learning model;   retrieve a second vector representation of a second entity that the user has not interacted with, wherein the second vector representation is generated using the initial deep-learning model;   determine a first similarity score between the first vector representation and a vector representation of the user;   determine a second similarity score between the second vector representation and the vector representation of the user;   upon determining that the second similarity score is greater than the first similarity score, update, using the initial deep-learning model, the second vector representation; and   generate an updated deep-learning model based on the initial deep-learning model and the updated second vector representation.   
     
     
         20 . A system comprising: one or more processors; and a memory coupled to the processors comprising instructions executable by the processors, the processors operable when executing the instructions to:
 retrieve a first vector representation of a first entity that a user has interacted with, wherein the first vector representation is generated using an initial deep-learning model;   retrieve a second vector representation of a second entity that the user has not interacted with, wherein the second vector representation is generated using the initial deep-learning model;   determine a first similarity score between the first vector representation and a vector representation of the user;   determine a second similarity score between the second vector representation and the vector representation of the user;   upon determining that the second similarity score is greater than the first similarity score, update, using the initial deep-learning model, the second vector representation; and   generate an updated deep-learning model based on the initial deep-learning model and the updated second vector representation.