An informed framework for training classifiers from social media

Abstract

Extracting information from social media has become a major focus of companies and researchers in recent years. Aside from the study of the social aspects, it has also been found feasible to exploit the collaborative strength of crowds to help solve classical machine learning problems like object recognition. In this work, we focus on the generally underappreciated problem of building effective datasets for training classifiers by automatically assembling data from social media. We detail some of the challenges of this approach and outline a framework that uses expanded search queries to retrieve more qualified data. In particular, we concentrate on collaboratively tagged media on the social platform Flickr, and on the problem of image classification to evaluate our approach. Finally, we describe a novel entropy-based method to incorporate an information-theoretic principle to guide our framework. Experimental validation against well-known public datasets shows the viability of this approach and marks an improvement over the state of the art in terms of simplicity and performance.