We investigate parameter-based and distributionbased approaches to regularizing the generative, similarity-based classifier called local similarity discriminant analysis classifier (local SDA). We argue that regularizing distributions rather than parameters can both increase the model flexibility and decrease estimation variance while retaining the conceptual underpinnings of the local SDA classifier. Experiments with four benchmark similarity-based classification datasets show that the proposed regularization significantly improves classification performance compared to the local SDA classifier, and the distributionbased approach improves performance more consistently than the parameter-based approaches. Also, regularized local SDA can perform significantly better than similarity-based SVM classifiers, particularly on sparse and highly nonmetric similarities.Keywords-local similarity discriminant analysis; regularized local similarity discriminant analysis;
I. SIMILARITY-BASED CLASSIFICATIONSimilarity-based classifiers learn from a set of pairwise training similarities, training class labels, and from the similarities between a test sample and the training samples [1]. Similarity-based classifiers are independent of a chosen similarity measure, which is usually problem-dependent and can subsume complex relationships between complex, heterogeneous samples. In this paper, we focus on the problem of designing generative classifiers for similarity-based learning. Here, the goal is to create class-conditional probabilistic models of the given similarities. Generative similarity-based classifiers differ from the standard metric-based generative classifiers, such as quadratic discriminant analysis and Gaussian mixture models, because the modeled quantity is the pairwise similarities between the samples rather than the numerical feature vectors that describe the samples. Producing class probabilities is important in many practical systems where there may be skewed class priors or asymmetric misclassification costs, or where probabilities are required as an input to the next component in the system or to fuse with probabilistic information about the class label derived from other sources.Recently, an effective generative classifier for similaritybased learning called similarity discriminant analysis (SDA) Work supported by the U.S. Office of Naval Research and a local version (local SDA) were proposed [2], [3]. We review local SDA in Section 3, and discuss how this classifier can fail. In Section 4, we follow our analysis with a discussion of several regularization strategies for local SDA and with the main contribution of this paper: that appropriate regularization can both make the SDA model more flexible and lower the estimation variance. Experiments in Section 5 show that the proposed regularized local SDA improves on local SDA and can outperform other state-of-the-art similaritybased classifiers.Previous research on generative classifiers for similaritybased learning treated the n-vector of similarities between any ...