Noise and uncertainty management in intelligent data modeling

Abstract

The management of uncertain and noisy data plays an important role in many problem solving tasks. One traditional approach is to quantify the magnitude of noise or uncertainty in the data and to take this information into account when using this type of data for different purposes. In this paper we propose an alternative way of handling uncertain and noisy data. In particular, noise in the data is positively identified and deleted so that quality data can be obtained. Using the assumption that interesting properties in data are more stable than the noise, we propose a general strategy which involves machine learning from data and domain knowledge. This strategy has been shown to provide a satisfactory way of locating and rejecting noise in large quantities of visual field test data, crucial for the diagnosis of a variety of blinding diseases.