Deep semantic content selection

A non-transitory computer readable storage medium storing instructions, the instructions when executed by a processor cause the processor to:

• determine a first user latent vector for a user and a first object latent vector for a content item based on a first object associated with the content item, the first object having a first object type, wherein the first user latent vector and the first object latent vector are in a first latent space and determined using a first model trained based on co-occurrence of interactions of a first set of users with a first set of content objects having the first object type;
• determine a second user latent vector and a second object latent vector for the content item based on a second object associated with the content item, the second object having a second object type, wherein the second user latent vector and the second object latent vector are in a second latent space and determined using a second model trained based on co-occurrence of interactions of a second set of users with a second set of content objects having the second object type, the second model different than the first model;
• determine a third user latent vector for the user and a third object latent vector for the content item based on a third object associated with the content item, the third object having a third object type, wherein the third user latent vector and the third object latent vector are in a third latent space and determined using a third model trained based on co-occurrence of interactions of a third set of users with a third set of content objects having the third object type, the third model different than the first model and the second model;
• determine a content item latent vector based on the first object latent vector, the second object latent vector, and the third object latent vector;
• determine a user latent vector based on the first user latent vector, the second user latent vector, and the third user latent vector;
• and
• determine, using a machine-learned model trained based on past interactions of a plurality of users and a plurality of content items, a score indicative of a likelihood of the user interacting with the content item, the score based on the content item latent vector and the user latent vector.

A system comprising:

• one or more processors;
• and
• a non-transitory computer readable storage medium storing instructions, the instructions when executed by a processor cause the processor to:
• determine a first user latent vector for a user and a first object latent vector for a content item based on a first object associated with the content item, the first object having a first object type, wherein the first user latent vector and the first object latent vector are in a first latent space and determined using a first model trained based on co-occurrence of interactions of a first set of users with a first set of content objects having the first object type;
• determine a second user latent vector and a second object latent vector for the content item based on a second object associated with the content item, the second object having a second object type, wherein the second user latent vector and the second object latent vector are in a second latent space and determined using a second model trained based on co-occurrence of interactions of a second set of users with a second set of content objects having the second object type, the second model different than the first model;
• determine a third user latent vector for the user and a third object latent vector for the content item based on a third object associated with the content item, the third object having a third object type, wherein the third user latent vector and the third object latent vector are in a third latent space and determined using a third model trained based on co-occurrence of interactions of a third set of users with a third set of content objects having the third object type, the third model different than the first model and the second model;
• determine a content item latent vector based on the first object latent vector, the second object latent vector, and the third object latent vector;
• determine a user latent vector based on the first user latent vector, the second user latent vector, and the third user latent vector;
• and
• determine, using a machine-learned model trained based on past interactions of a plurality of users and a plurality of content items, a score indicative of a likelihood of the user interacting with the content item, the score based on the content item latent vector and the user latent vector.