Research in interactive machine learning has shown the effectiveness of live human interaction with machine learning algorithms in many applications. Metric learning is a common type of algorithm employed in this context, using feedback from users to learn a distance metric over the data that encapsulares their own understanding. Less progress has been made on helping users decide which data to ex... read moreamine for potential feedback. Systems may make suggestions for grouping items, or may propose constraints to the user, generally by focusing on fixing areas of uncertainty in the model. For this work in progress, we propose an active learning approach, aimed at an interactive machine learning context, that tries to minimize user effort by directly estimating the amount of change that potential inputs will have on the model and querying users appropriately. With EigenSense, we use eigenvector sensitivity in the pairwise distance matrix induced by a distance metric over the data to estimate how much a given user input might affect the metric. We evaluate the techinique by comparing the output points it proposes for user consideration against what an oracle would like to choose as inputs.read less
Brown, Eli T. and Remco Chang. "EigenSense: Saving User Effort with Active Metric Learning." Paper presented at Workshop on on Interactive Data Exploration and Analytics (IDEA), KDD2014, New York, New York, August 24-27, 2014.