Dynamic Clustering-Based Estimation of Missing Values in Mixed Type Data

Abstract

Abstract. The appropriate choice of a method for imputation of missing data becomes especially important when the fraction of missing values is large and the data are of mixed type. The proposed dynamic clustering imputation (DCI) algorithm relies on similarity information from shared neighbors, where mixed type variables are considered together. When evaluated on a public social science dataset of 46,043 mixed type instances with up to 33% missing values, DCI resulted in more than 20% improved imputation accuracy over Multiple Imputation, Predictive Mean Matching, Linear and Multilevel Regression, and Mean Mode Replacement methods. Data imputed by 6 methods were used for test of NB-Tree, Random Subset Selection and Neural Network-based classification models. In our experiments classification accuracy obtained using DCI-preprocessed data was a lot better than when relying on alternative imputation methods for data preprocessing.

Handling Missing Values via a Neural Selective Input Model

“…These results seem intuitive since in principle multiple imputation works better when the proportion of MVs is smaller, in which case more data are available for validating the estimates inferred. These results assume particular relevance, if we consider that the appropriate choice of the method for handling MVs is especially important when the fraction of MVs is large [18].…”

“…Since many algorithms cannot directly handle MVs, a common practice is to rely on data pre-processing techniques. Usually, this is accomplished by using imputation or simply by removing instances (case deletion) and/or features containing MVs [5,3,11,12,17,18,1]. A review of the methods and techniques to deal with this problem, including a comparison of some well-known approaches, can be found in Laencina et al [5].…”

“…This is a very conservative strategy which guarantees that errors are not introduced in the dataset [6]. However, it is not applicable when the MVs cover a large fraction of the instances, or when their presence in essential attributes is large [18,6]. In some cases, MVs represent only a small fraction of the data but they spread throughout a large number of instances, rendering the option of case deletion inviable.…”

“…The results of the NSIM were compared with single imputation and multiple imputation methods. Multiple imputation is considered one of the most powerful approaches for estimating MVs [18]. For single imputation the version 3.7.5 of the Weka software package was used [26].…”