An Application of Divergence Estimation to Projection Retrieval for Semi-supervised Classication and Clustering
Madalina Fiterau, Artur W. Dubrawski · 2013
Decision-support systems enable human users to handle tasks such as classication or clustering. Often, this expert intervention involves simply the validation of an outcome produced by an automated system. In such cases, the decision process must be made transparent and comprehensible. We formulate the projection retrieval problem as an optimization using a divergence-based objective specic to the learning task. The framework presented in the paper uses consistent divergence estimators to extract a set of lowdimensional projections which jointly form a solution to the learning task. The method works under the assumption that each projection addresses a dierent subspace of the feature space and that it is possible to learn a function selecting the appropriate projection for a given test point. Experiments show that the method recovers the underlying structure of the data and provides low-dimensional views that aid expert assessment.