Systems and methods for estimating sender similarity based on user labels

Abstract

A method assigns categories to electronic messages. Users assign labels to messages received from multiple senders. The user-assigned labels are aggregated into a co-occurrence matrix whose axes correspond to the plurality of senders. Each matrix entry measures the extent to which users have assigned identical user-assigned labels to the pair of senders. Using the co-occurrence matrix, a latent vector is computed for each sender. A user re-categorizes a message received from a first sender, assigning the message to a first category. One or more second senders are identified that are similar to the first sender. Similarity of senders is measured by proximity of the latent vectors corresponding to the senders. One or more second messages are identified from the second senders, and the user is presented with the suggestion to reclassify the second messages under the first category. Upon user confirmation, the second messages are reclassified.