Predicting Missing Attribute Values based on Frequent Itemset and RSFit

How to process missing attribute values is an important data preprocessing problem in data mining and knowledge discovery tasks. A commonly-used and naive solution to process data with missing attribute values is to ignore the instances which contain missing attribute values. This method may neglect important information within the data and a significant amount of data could be easily discarded. Some methods, such as assigning the most common values or assigning an average value to the missing attribute, make good use of all the available data. However the assigned value may not come from the information which the data originally derived from, thus noise is brought to the data. We introduce an integrated approach ItemRSFit to effectively predict missing attribute values by combining frequent itemset and RSFit approaches together. Frequent itemset is generated from the association rules algorithm and it displays the correlations between different items in a transaction data set. Using frequent itemset as a knowledge base to predict missing attribute values is shown to have a high prediction accuracy. However this approach alone cannot predict all the existing missing attributes. RSFit is a newly developed approach to predict missing attribute values based on the similarities of attribute-value pairs by only considering attributes contained in the core or the reduct of the data set. The RSFit approach provides a faster prediction and can be used for the cases that are not covered by the itemset approach. Empirical studies on UCI data sets and a real world data set demonstrate a significant increase of predicting accuracy obtained from this new integrated approach.