Training generative models on feature-level relationships

A computing device, comprising:

• one or more processors;
• and
• one or more memories storing computer-executable instructions that, when executed by the one or more processors, cause the one or more processors to:
• receive a first image associated with a first area of an article representing a transfer of at least one resource from a first entity to a second entity, wherein the first image represents a first user-defined feature of the article;
• receive a second image associated with a second area of the article, wherein the second image represents a second user-defined feature of the article;
• provide, to a generative model, the first image and the second image;
• receive, from the generative model, a first embedding associated with the first image and a second embedding associated with the second image;
• generate a combined embedding by combining the first embedding and the second embedding;
• determine a vector distance metric based at least in part on a vector distance comparison between the combined embedding and a plurality of prior embeddings stored in a prior embedding datastore;
• generate a training label indicating a match or mismatch based at least in part on the vector distance metric exceeding a predefined threshold;
• assign a sample weight to the training label that changes as a function of an age of a corresponding prior article from which a prior embedding of the plurality of prior embeddings was generated;
• and
• produce a trained generative model by training the generative model by modifying weights based at least in part on an error between a first generated output of the generative model and the training label.

A non-transitory computer-readable storage medium storing computer-executable instructions that, when executed with one or more processors of a computing device, cause the one or more processors to:

• receive a first image associated with a first area of an article representing a transfer of at least one resource from a first entity to a second entity, wherein the first image represents a first user-defined feature of the article;
• receive a second image associated with a second area of the article, wherein the second image represents a second user-defined feature of the article;
• provide, to a generative model, the first image and the second image;
• receive, from the generative model, a first embedding associated with the first image and a second embedding associated with the second image;
• generate a combined embedding by combining the first embedding and the second embedding;
• determine a vector distance metric based at least in part on a vector distance comparison between the combined embedding and a plurality of prior embeddings stored in a prior embedding datastore;
• generate a training label indicating a match or mismatch based at least in part on the vector distance metric exceeding a predefined threshold;
• assign a sample weight to the training label that changes as a function of an age of a corresponding prior article from which a prior embedding of the plurality of prior embeddings was generated;
• and
• produce a trained generative model by training the generative model by modifying weights based at least in part on an error between a first generated output of the generative model and the training label.