Building a training dataset for classification under a cost limitation

A necessary preprocessing of document classification is to label some documents so that a classifier can be built based on which the remaining documents can be classified. Because each document differs in length and complexity, the cost of labeling each document is different. The purpose of this paper is to consider how to select a subset of documents for labeling with a limited budget so that the total cost of the spending does not exceed the budget limit, while at the same time building a classifier with the best classification results.

In this paper, a framework is proposed to select the instances for labeling that integrate two clustering algorithms and two centroid selection methods. From the selected and labeled instances, five different classifiers were constructed with good classification accuracy to prove the superiority of the selected instances.

Experimental results show that this method can establish a training data set containing the most suitable data under the premise of considering the cost constraints. The data set considers both “data representativeness” and “data selection cost,” so that the training data labeled by experts can effectively establish a classifier with high accuracy.

No previous research has considered how to establish a training set with a cost limit when each document has a distinct labeling cost. This paper is the first attempt to resolve this issue.