Missing value estimation based on dynamic attribute selection

Abstract

Raw Data used in data mining often contain missing information, which inevitably degrades the quality of the derived knowledge. In this paper, a new method of guessing missing attribute values is suggested. This method selects attributes one by one using attribute group mutual information calculated by flattening the already selected attributes. As each new attribute is added, its missing values are filled up by generating a decision tree, and the previously filled up missing values are naturally utilized. This ordered estimation of missing values is compared with some conventional methods including Lobo's ordered estimation which uses static ranking of attributes. Experimental results show that this method generates good recognition ratios in almost all domains with many missing values.