Abstract:
Minimizing empirical risk, i.e., training error, means that a classifier should have a high classification accuracy on label-known samples/training samples, while minimizing Vapnik-Chemonenkis (VC) dimension or complexity means that a classifier should have a low prediction error on label-unknown samples/test samples. Since structural risk includes empirical risk and complexity, it can be utilized to measure the effectiveness of a classifier. The structural risk minimization (SRM) means that a classifier should have better classification accuracy on training samples and better prediction performance on test samples simultaneously. Since a global space can be divided into several local ones, the minimization problem of structural risk has global and local minimization. In recent years, multi-view classifiers have been developed to process multi-view data sets composed of the samples with multiple views. Each view represents the information of data set from a certain field. However, traditional multi-view classifiers cannot minimize the structural risk fully due to ignoring the differences of views and features. Aiming at the above shortcoming, this paper proposes a weight-based multi-view global and local structural risk minimization classifier. This proposed classifier utilizes the weight of features and the weight of views so that the classifier can attain the consistence with the distribution of data set. It is shown from experiment results on three multi-view data sets, Mfeat, Reuters, and Corel that by introducing the weights of features and views, the corresponding multi-view classifier has better classification performance. Hence, the proposed method can enhance the performances of classifiers and effectively minimize structural risk.