Object identification based on long-term user behavior and short-term interest

A computer-implemented method, comprising:

• generating a plurality of object embeddings, each object embedding representative of an object of a first plurality of objects, each object embedding including continuous data representative of attributes of the object represented by the object embedding;
• receiving a text-based request for an object of interest to a user;
• obtaining, from a user profile embeddings data store, a user profile embedding including first continuous data representative of a long-term behavior of the user;
• determining a second plurality of objects of the first plurality of objects with which the user has interacted during a defined period of time;
• combining the continuous data of each respective object embedding of the second plurality of objects and the first continuous data of the user profile embedding to generate second continuous data that is included in a user embedding;
• comparing the second continuous data of the user embedding with continuous data of object embeddings of a third plurality of objects included in the first plurality of objects to generate, for each comparison, a first similarity score;
• determining, for each of the third plurality of objects having highest first similarity scores, a second similarity score indicative of a similarity between the object and an input embedding representative of the text-based request;
• and
• returning at least a first object of the third plurality of objects having a highest second similarity score as an object of potential interest to the user.

A non-transitory computer-readable storage medium storing instructions that, when executed by at least one processor of a computing system, cause the computing system to at least:

• generate a first plurality of object embeddings for a first plurality of objects, wherein the first plurality of object embeddings include continuous data representative of a plurality of attributes corresponding to the first plurality of objects;
• obtain a user profile embedding of a first user, the user profile embedding including first continuous data representative of a long-term behavior of the first user;
• receive a text-based request for content from the first user;
• and
• in response to receipt of the text-based request:
• determine a second plurality of objects from the first plurality of objects with which the first user has interacted during a defined period of time;
• combine the continuous data of the object embeddings of each object of the second plurality of objects and the continuous data of the user profile embedding to generate second continuous data that is included in a user embedding of the first user;
• compare the second continuous data of the user embedding with continuous data of object embeddings of a third plurality of objects to generate, for each comparison, a similarity score for each object of the third plurality of objects, wherein the third plurality of objects are included in the first plurality of objects;
• and
• return a first object of the third plurality of objects having a highest similarity score as an object of potential interest to the user.