coLabel: Active Learning for Internet Image Tags

This paper considers the problem of machine learning image classification from social media website data. People on socialmedia websites produce behemoth amounts of data for machine learning. However, the wide-range of image types and labeling confusion makes it difficult to train high-quality classifiers. This paper proposes an approach to automatically identify mistagged images using ensembles of learning algorithms. We develop a number of ranking algorithms, and show their correlation with a mislabeling detected using a small website experiment.