Feature subspace isolation and disentanglement in merchant embeddings

A system for making merchant recommendations for a target user, comprising:

• a transaction database storing payment transaction records;
• a hardware processor having access to the transaction database;
• and
• a software component containing program instructions, which when executed by the hardware processor cause the hardware processor to:
• receive raw merchant embeddings generated from the payment transaction records without any rating data, review data, or browsing session data, wherein the raw merchant embeddings represent merchants in different locations and include a plurality of embedded features, including a location feature and a price feature that are entangled in an embedding space such that the location feature causes the raw merchant embeddings for the merchants in different locations to be separated in the embedding space, the raw merchant embeddings originating from both a source region most frequented by the target user, and a target region for which merchant recommendations are made for the target user;
• use transaction metadata associated with the payment transaction records to determine a subspace of the location feature within the embedding space by:
• representing the raw merchant embeddings originating from the source region by a first vector space V1=v1, v2 . . . vn in a source embedding space, and representing the raw merchant embeddings originating from the target region by a second vector space V2=V1, v2 . . . vm in a target embedding space;
• selecting a portion of the raw merchant embeddings from the first vector space V1 as anchor points in the source embedding space, and selecting a portion of the raw merchant embeddings from the second vector space V2 as anchor points in the target embedding space;
• and
• determining a direction of bias (b) between the anchor points in the first vector space V1 and the second vector space V2, wherein the direction of bias (b) defines the subspace of the location feature;
• remove the subspace of the location feature from the embedding space using a linear transformation process to create modified merchant embeddings in which remaining embedded features, including the price feature, a popularity/frequency feature, and a cuisine type feature, are merged and aligned within the embedding space so that the modified merchant embeddings representing the merchants in the different locations occupy a similar region of the embedding space;
• generate a list of merchant rankings based on the modified merchant embeddings, past preferences of the target user determined using raw user embeddings, and the target region;
• and
• provide the list of merchant rankings to the target user.

A non-transitory computer-readable medium containing program instructions for making merchant recommendations for a target user, the instructions comprising:

• receiving, by one or more processors, raw merchant embeddings generated from payment transaction records without any rating or review data or browsing session data, wherein the raw merchant embeddings originate from a source region most frequented by the target user, and a target region for which merchant recommendations are made for the target user, and wherein the raw merchant embeddings represent merchants in different locations and include a plurality of embedded features entangled in an embedding space, the embedded features including a location feature, a price feature and a cuisine type feature, such that the location feature causes the raw merchant embeddings for the merchants in different locations to be separated in the embedding space;
• using, by the one or more processors, transaction metadata associated with the payment transaction records to determine a subspace of the location feature within the embedding space by:
• determining a direction of bias (b) between anchor points in a first vector space V1=v1, v2 . . . vn representing the raw merchant embeddings originating from the source region and anchor points in a second vector space V2=v1, v2 . . . vm representing the raw merchant embeddings originating from the target region, wherein the direction of bias (b) defines the subspace of the location feature and is calculated as:
• b
• =
• mean
• ⁡
• (
• v
• ⁢
• ⁢
• 1
• )
• -
• mean
• ⁡
• (
• v
• ⁢
• ⁢
• 2
• )
• 
• mean
• ⁡
• (
• v
• ⁢
• ⁢
• 1
• )
• -
• mean
• ⁡
• (
• v
• ⁢
• ⁢
• 2
• )
• 
• ;
• removing, by the one or more processors, the subspace of the location feature from the embedding space using a linear transformation process to create modified merchant embeddings in which remaining embedded features are merged and aligned within the embedding space by calculating v′=v−<v, b>b, where v=anchor points from V1 or V2, so that the modified merchant embeddings representing the merchants in the different locations occupy a similar region of the embedding space;
• generating, by the one or more processors, a list of merchant rankings based on the modified merchant embeddings, past preferences of the target user determined using raw user embeddings, and the target region;
• and
• providing, by the one or more processors, the list of merchant rankings to the target user.