Using Instance Typicality to Build Compact and Accurate Neural Network Classifiers

Classification is one of the commonly used tasks in data mining. Classification accuracy, training time and storage requirement are some of the important issues in the design of the classifiers. Techniques, such as bagging, boosting and ensembles exist to improve accuracy of classifiers. Feature selection and instance selection algorithms are often used to reduce training times and storage requirements of the models. Instance-based classifiers use concept of instance typicality to improve the accuracy of the classifiers by selecting most typical instances from the training set by computing typicality scores explicitly. The cost of this computation is high, especially for large datasets. Neural network classifiers, however, need not compute typicality scores explicitly as the values at the outputs of neurons in the output layer represent typicality scores. This paper presents a novel wrapper model to construct compact classifiers using an algorithm that selects only prototype instances. Experimental results show that the model results into significant reduction in storage space without affecting the classification accuracy in certain problem domains.