|
Abstract:
|
Semi-supervised classification---training first on labeled observations and subsequently updating the classifier based on unlabeled observations---can yield improved performance compared to the classifier based on only the labeled observations. However, use of the unlabeled observations may degrade the performance when the model is misspecified. Many factors affect the semi-supervised performance, including training data, model specification, estimation method, and the classifier itself. For concreteness, we consider maximum likelihood estimation in Gaussian mixture models and the Bayes plug-in classifier, due to their ubiquitousness and tractability. In this setting, we examine the effect of model misspecification on semi-supervised classification performance and shed some light on when and why performance degradation obtains.
|
