Impact of outlier detection on neural networks based property value prediction

Abstract

Detecting outliers is an important step in data mining. Outliers not only hamper data quality but also affect the output in case of prediction models. Prediction tools like Neural Networks (NN) need outlier free dataset in order to achieve better generalization of the network as errors in the dataset hinder the modelling process and produce misleading results. Thus, range of the dataset needs to be curbed in order to make it fit for generating better prediction results. However, outlier detection faces one difficulty. There is no standard framework for the treatment of outliers found in the literature. The present study is an effort to identify the most suited outlier detection method for a specific problem, which deals with the use of NN for prediction of real property value. 3094 cases of property sale instances are presented to various univariate outlier detection methods like Tukey’s method, Standard Deviation (SD) method, median method, Z score method, MAD method, modified Z score method, etc. The datasets prepared after removing outliers marked for respective methods are used for prediction using NN. Comparison of results show that the median method is the best-suited outlier detection method for the present study.