Solving missing data problems in data quality with deep learning method: An application with generative adversial networks

Revolutionary developments in the field of big data analysis and machine learning algorithms have changed the business strategies of industries such as banking, financial services, asset management, food and beverage and ecommerce. Data-driven decision-making strategies of these enterprises increased their competitiveness, causing them to face some new problems. One of the most common problems that businesses face when using data is missing data in data sets. Missing data problem is one of the important factors affecting data quality. The aim of the research is to choose the appropriate method to solve the missing data problems affecting the data quality and to create a guide for the wine producing enterprises that they can apply against the missing data problems. For this purpose, the application was made on the new data set obtained by creating missing values in the data set named wine quality. In this context, the differentiable generative model was used experimentally in completing missing data. The generative adversial imputation networks (GAIN) represent the class of differentiable generative models. The performance of the GAIN model was compared with traditional methods, Decision trees, multiple imputation, expectation maximization. In addition, the wasserstein generative advsersial imputation networks(WGAIN) algorithm, which was developed with the wasserstein setting of the GAIN model, was introduced. The evaluation of the algorithm was made by calculating the differences between the original data set and the completed data set values. The RMSE evaluation criterion was used for this. Empirical findings, GAIN algorithm, especially WGAIN algorithm has been found to show the most successful performance in each missing mechanism and missing rates from 10% to 50%. The solution stages of the missing data problem in the wine producing enterprise scale are explained and it is stated that the most suitable method, WGAIN algorithm should be preferred. Analysis is important since GAIN is included among the missing data analysis methods, which is rarely found in Turkish sources.