Composite embedding systems and methods for multi-level granularity similarity relevance scoring

A non-transitory computer readable medium having instructions stored thereon, wherein the instructions, when executed by a processor cause a device to perform operations comprising:

• receiving, from a user system over a network via a network interface, a target string;
• obtaining from an active database a set of common strings, wherein each string in the set of common strings represents a number of associated frequently used search queries that contains the respective string, and the set of common strings is updated periodically;
• comparing the target string to the set of common strings;
• in accordance with a determination that the target string does not match any string in the set of common strings, generating an anchor embedding for the target string, wherein the anchor embedding is a composite embedding generated by concatenating:
• at least a first initial embedding with a first number of dimensions, and a second initial embedding with a second number of dimensions, wherein the first number is different from the second number;
• obtaining an item embedding for each item in a pool of items, wherein:
• each item in the pool of items is associated with at least one string in the set of common strings;
• a first plurality of item embeddings for the pool of items are pre-generated composite embeddings;
• a second plurality of the item embeddings for the pool of items are generated by:
• obtaining, from the database, a first set of composite embedding parameters, wherein the first set of composite embedding parameters identify the first initial item embedding with the first number of dimensions and associated first weights;
• obtaining, from the database, a second set of composite embedding parameters, wherein the second set of composite embedding parameters identify the second initial item embedding with the second number of dimensions and associated second weights;
• based on the first set of composite embedding parameters, generating a respective weighted first initial item embedding with the first number of dimensions by applying the first weights to the respective first initial item embedding;
• based on the second set of composite embedding parameters, generating a respective weighted second initial item embedding with the second number of dimensions by applying the second weights to the respective second initial item embedding;
• and
• concatenating the respective weighted first initial item embedding with the first number of dimensions and the respective weighted second initial item embedding with the second number of dimensions, wherein the respective weighted first initial item embedding and the respective weighted second initial item embedding are generated for the item, wherein;
• the item embedding for each item in the pool of items is the same dimensions as the anchor embedding;
• generating a similarity score for each item in the pool of items based on a single comparison of the item embedding to the anchor embedding;
• selecting a set of items from the pool of items based on the generated similarity scores, wherein the set of items includes a predetermined number of items in the pool of items having a highest similarity score;
• and
• transmitting, over the network via the network interface, the selected set of items to populate elements of a user interface.

A method by at least one processor, the method comprising:

• receiving, from a user system over a network via a network interface, a target string;
• obtaining from an active database a set of common strings, wherein each string in the set of common strings represents a number of associated frequently used search queries that contain the respective string, and the set of common strings is updated periodically such as one of every hour or every day;
• comparing the target string to the set of common strings;
• in accordance with a determination that the target string does not match any string in the set of common strings, generating an anchor embedding for the target string, wherein the anchor embedding is a composite embedding generated by concatenating:
• at least a first normalized initial embedding with a first number of dimensions;
• and
• a second normalized initial embedding with a second number of dimensions, wherein the first number is different from the second number;
• obtaining an item embedding for each item in a pool of items, wherein:
• each item in the pool of items is associated with at least one string in the set of common strings;
• a first plurality of item embeddings for the pool of items are pre-generated composite embeddings;
• a second plurality of the item embeddings for the pool of items is generated by:
• obtaining, from the database, a first set of composite embedding parameters, wherein the first set of composite embedding parameters identify the first initial item embedding with the first number of dimensions and associated first weights;
• obtaining, from the database, a second set of composite embedding parameters, wherein the second set of composite embedding parameters identify the second initial item embedding with the second number of dimensions and associated second weights;
• based on the first set of composite embedding parameters, generating a respective weighted first initial item embedding with the first number of dimensions by applying the first weights to the respective first initial item embedding;
• based on the second set of composite embedding parameters, generating a respective weighted second initial item embedding with the second number of dimensions by applying the second weights to the respective second initial item embedding;
• and
• concatenating the respective weighted first normalized initial item embedding with the first number of dimensions and the respective weighted second normalized initial item embedding with the second number of dimensions, wherein the respective weighted first normalized initial item embedding and the respective weighted second normalized initial item embedding are generated for the item, wherein
• the item embedding for each item in the pool of items is the same dimensions as the anchor embedding;
• generating, based on a single comparison, a cosine similarity score between each item embedding and the anchor embedding;
• selecting a set of items from the pool of items based on the generated cosine similarity scores, wherein the set of items includes a predetermined number of items having a highest cosine similarity;
• and
• transmitting, over the network via the network interface, the selected set of items to populate elements of a user interface.