Geo-encoded embeddings

One or more computer-readable non-transitory storage media embodying software that is operable when executed to:

• generate embeddings for social-networking entities by training the embeddings using a training algorithm, wherein an embedding corresponding to an entity represents a point in a d-dimensional embedding space, and wherein contextually similar entities are represented by embeddings close to each other in the embedding space;
• identify a subset of entities having one or more common attributes that is not encoded in the generated embeddings;
• encode, for each entity in the subset, values of the one or more common attributes into a j-dimensional additional embedding;
• create, for each entity in the subset, a (d+j)-dimensional embedding by concatenating the generated d-dimensional embedding with the j-dimensional additional embedding;
• detect a need to identify entities similar to a reference entity that is a member of the subset;
• compute k-nearest neighbors of an embedding corresponding to the reference entity in the (d+j)-dimensional embedding space;
• identify entities corresponding to the computed k-nearest neighbors;
• and
• provide information regarding the corresponding entities.

A system comprising:

• one or more processors;
• and
• one or more computer-readable non-transitory storage media coupled to one or more of the processors and comprising instructions operable when executed by one or more of the processors to cause the system to:
• generate embeddings for social-networking entities by training the embeddings using a training algorithm, wherein an embedding corresponding to an entity represents a point in a d-dimensional embedding space, and wherein contextually similar entities are represented by embeddings close to each other in the embedding space;
• identify a subset of entities having one or more common attributes that is not encoded in the generated embeddings;
• encode, for each entity in the subset, values of the one or more common attributes into a j-dimensional additional embedding;
• create, for each entity in the subset, a (d+j)-dimensional embedding by concatenating the generated d-dimensional embedding with the j-dimensional additional embedding;
• detect a need to identify entities similar to a reference entity that is a member of the subset;
• compute k-nearest neighbors of an embedding corresponding to the reference entity in the (d+j)-dimensional embedding space;
• identify entities corresponding to the computed k-nearest neighbors;
• and
• provide information regarding the corresponding entities.