Improving classification performance on real data through imputation

The applicability of learning methods to raw data coming from different areas of human activity is one of the main concerns in data mining research today. This paper emphasizes the need for a sound preprocessing method to improve the quality of the learning process through data imputation. Three classification methods we have previously developed are presented, with a focus on their evaluations. The results prove their increased performance on benchmark data, when compared to similar approaches. Although on real-world data improvements have been observed as well, the case study presented here has revealed the need for a reliable preprocessing method, to enhance the performance of the methods on real, incomplete data. We have carried out preliminary evaluations, on benchmark data, with a new imputation method, based on an ensemble of neural networks.