Text-based similarity system for cold start recommendations

A method of generating cold-start recommendations based on title and description relationships, the method comprising:

• creating a set of training samples, each training sample comprising a title and a description, from a domain-specific corpus associated with a plurality of candidate items, the domain-specific corpus comprising a title and a description for each candidate item in the plurality of candidate items;
• creating a set of real title-description pairs and a set of fake title-description pairs within the set of training samples, each real title-description pair in the set of real title-description pairs comprising a title and a description of a same item and each fake title-description pair in the set of fake title-description pairs comprising a title from a first item and a description from a second different item;
• training a language model, the language model comprising a masking component and a reconstruction component, to distinguish between real title-description pairs and fake title-description pairs, wherein:
• the masking component randomly masks a pre-determined percentage of words in each training sample;
• and
• the reconstruction component reconstructs the masked words,
• and wherein the trained language model is used to generate, using the trained language model, a set of scores for each candidate item in the plurality of candidate items and a seed item, the set of scores comprising a set of similarity scores and a set of classification scores associated with combinations of the seed item title, the seed item description, each candidate item title and each candidate item description;
• selecting a set of recommended items from the set of candidate items based on the set of scores to maximize item similarities;
• and
• outputting, via a user interface, the set of recommended items to at least one user.

One or more computer storage devices having computer-executable instructions stored thereon for generating cold-start recommendations based on title and description relationships, which, on execution by a computer, cause the computer to perform operations comprising:

• creating a set of training samples, each training sample comprising a title and a description, from a domain-specific corpus associated with a plurality of candidate items, the domain-specific corpus comprising a title and a description for each candidate item in the plurality of candidate items;
• creating a set of real title-description pairs and a set of fake title-description pairs within the set of training samples, each real title-description pair in the set of real title-description pairs comprising a title and a description of a same item and each fake title-description pair in the set of fake title-description pairs comprising a title from a first item and a description from a second different item;
• generating, using a language model, a set of scores for each candidate item in the plurality of candidate items and a seed item, the set of scores comprising a set of similarity scores and a set of classification scores associated with combinations of the seed item title, the seed item description, each candidate item title and each candidate item description, wherein the generating the set of scores comprises training the language model to distinguish between real title-description pairs and fake title-description pairs, wherein:
• the masking component randomly masks a pre-determined percentage of words in each training sample;
• and
• the reconstruction component reconstructs the masked words;
• selecting a set of recommended items from the set of candidate items based on the set of scores to maximize item similarities;
• and
• outputting, via a user interface, the set of recommended items to at least one user.