A simple and efficient incremental missing data imputation method for evolving neo-fuzzy network

Abstract

A great challenge in real-world applications driven which use data streams to solve forecast problems is handling missing data. Although there are methods to reduce the effects caused by this issue, most systems are not modeled in a preventive way to enable an adequate treatment of this type of occurrence. In this context, this paper introduces a new evolving fuzzy approach called evolving Neo-Fuzzy Neuron with Missing Data Procedure (eNFN-MDP), that handles single and multiple missing values on data samples. eNFN-MDP checks whether there are variables with missing values for each new sample. If one or more missing values are found, the estimated values are imputed. Then, the output is computed with all available values. Forecasting examples illustrate the usefulness of the approach. Experimental comparisons in Missing at Random and Missing Completely at Random in nonstationary environments are performed. The results of the eNFN-MDP are compared with state-of-the-art methods and models. Simulations results show that the eNFN-MDP achieves as high as or higher performance than the remaining evolving modeling methods. Therefore, the experimental results suggest the proposed approach as a simple and efficient alternative for data imputation in evolving modeling.