Content delivery based on corrective modeling techniques

A non-transitory computer-readable medium comprising computer program code, that, when executed by a processor, causes the processor to:

• receive, by an online system, a plurality of content items associated with a topic for delivery to a viewing user of the online system;
• for each of the plurality of content items:
• extract a first set of features, each feature in the first set based on one or more of user profile attributes of the viewing user or content item attributes of the content item, wherein one or more features in the first set of features are specific to the topic;
• apply the first set of features as input to a general model corresponding to the topic to generate a predicted output, wherein the predicted output represents a likelihood of the viewing user interacting with the content item;
• extract a second set of features, each feature in the second set based on one or more of user profile attributes of the viewing user or content item attributes of the content item, wherein one or more features in the second set of features are specific to a sub-topic of the topic and not included in the first set of features;
• apply the second set of features as input to a residual model corresponding to the sub-topic of the topic to generate a predicted residual value, wherein the predicted residual value is configured to adjust the predicted output of the general model based on the one or more features in the second set of features that are specific to the sub-topic;
• and
• generate a corrected predicted output by combining the predicted output and the predicted residual value;
• select a content item from the plurality of content items for delivery to a user of the online system based on the corrected predicted outputs of the plurality of content items;
• and
• send the selected content item to the user of the online system.

A method comprising:

• receiving general training data comprising training examples associated with a topic, each training example corresponding a content item and a user and comprising an output label that indicates whether an event has occurred;
• for each training example corresponding to a content item and a user:
• applying a first set of features of the training example as input to a general model corresponding to the topic to generate a predicted output that predicts whether the event has occurred, wherein one or more features in the first set of features are specific to the topic;
• determining a residual value by comparing the predicted output to the output label associated with the training example;
• applying a second set of features of the training example as input to a residual model corresponding to a sub-topic of the topic to generate a predicted residual value, wherein one or more features in the second set of features are specific to the sub-topic of the topic and not included in the first set of features;
• and
• determining a residual error by comparing the predicted residual value associated with the training example to the determined residual value associated with the training example, the residual error backpropagated to minimize the residual error.