Meta Platforms, Inc.
Iterative Classifier Training on Online Social Networks

Abstract:

In one embodiment, a method includes accessing a set of training objects associated with an object-classification, identifying, from comments associated with the training objects, by an initial object-classifier algorithm configured to classify objects as associated with an object-classification by comparing comments associated with each object to one or more features, a first set of features, each feature having a corresponding text expression and a feature score indicating a correlation value between the feature and the object-classification, adding the first set of features to the algorithm to generate a revised object-classifier algorithm, accessing a set of test objects, classifying one or more of the test objects as associated with the object-classification, identifying, from comments associated with the classified test objects, a second set of features having feature scores greater than a threshold feature score, and adding the second set of features to the revised algorithm to generate a final object-classifier algorithm.