Data-Driven Machine Learning Approach for Predicting Missing Values in Large Data Sets: A Comparison Study

Abstract

Pre-processing of large scale datasets in order to ensure data quality is a very important task in data mining. One of the serious threats to data quality is the lack of data collected during field experiments, which negatively affects the data quality. The missing data usually have significant effects in many real-life pattern classification scenarios, especially when it leads to biased parameter estimates but also disqualify for analysis purposes. The process of filling in the missing data based on other valid values of rest of the variables of a data set is known as the imputation process. In this paper, we present a new data-driven machine learning approach for imputing the missing data. Even though Machine Learning methods are used in order to impute missing data in the literature, it is difficult to decide on a single method to apply on a given data set for imputation. This is because imputation process is not considered as science but as art that focuses on choosing the best method with the least biased value. For this reason, we compare different machine learning methods, such as decision tree (C4.5), Bayesian network, clustering algorithm and artificial neural networks in this work. The comparison of the algorithms indicates that, for predicting categorical and numerical missing information in large survey data sets, clustering method is the most efficient out of the others methods found in literature. A hybrid method is introduced which combines unsupervised learning methods with supervised ones based on the missing ratio, for achieving a data imputation with higher accuracy. Additionally, some statistical imputation methods such as Mean\Mode, Hot-Deck have been applied emphasizing their limitations in large scale datasets in comparison to the machine learning methods. A comparison of all above mentioned methods, traditional statistical methods and machine learning methods has been made and conclusions are drawn for achieving data imputation with higher accuracy in data sets of large scale survey. Also, another objective of these experiments is to discover the effect of balancing the training data set in the performance of classifiers. All methods are tested to a real world data set, population and housing census.