Fast Data Acquisition in Cost-Sensitive Learning

摘要

Data acquisition is the first and one of the most important steps in many data mining applications. It is a time consuming and costly task. Acquiring an insufficient number of examples makes the learned model and future prediction inaccurate, while acquiring more examples than necessary wastes time and money. Thus it is very important to estimate the number examples needed for learning algorithms in machine learning. However, most previous learning algorithms learn from a given and fixed set of examples. To our knowledge, little previous work in machine learning can dynamically acquire examples as it learns, and decide the ideal number of examples needed. In this paper, we propose a simple on-line framework for fast data acquisition (FDA). FDA is an extrapolation method that estimates the number of examples needed in each acquisition and acquire them simultaneously. Comparing to the naive step-by-step data acquisition strategy, FDA reduces significantly the number of times of data acquisition and model building. This would significantly reduce the total cost of misclassification, data acquisition arrangement, computation, and examples acquired costs.