Fig. 6From: Task-agnostic representation learning of multimodal twitter data for downstream applicationsBot detection task. Component embeddings from a tweet are averaged; our CNN method (top right) uses a matrix of vectors from the same user as input to a CNN classifier to predict the user’s label, while our kNN method (bottom right) averages a user’s averaged embeddings and predicts the user’s label based on the k nearest usersBack to article page