Automatic Data Imputation in time seies precessin using neural networks for industry and medical datasets

JUAN PORTA; MARTÍN DOMÍNGUEZ; FRANCISCO TAMARIT

Lima

Conferencia; SIMBig 2021 - Springer CCIS; 2021

Time series classification and regression techniques help solveproblems in many knowledge areas, including medicine, electronics, in-dustry, and even music. When we apply them to real-life issues, a com-mon obstacle is the lack of data in intervals within a time series. Usually,to solve it, the missing data is populated with information highly de-pendent on available datasets, which requires prior analysis. This paperaddresses the problem in a novel way, automatically filling the missingdata using a mixture of techniques and letting the prediction model de-cide which filling is better. We tested our approach for classification inindustrial and medical datasets and for regression, we used a datasetcointaining COVID-19 information.Our results are very competitive, and our approach improves the state-of-the-art models. We obtain better performance in all the experimentsfor the selected quality measures. Most importantly, the improvement ismore statistically significant when the amount of missing data is higher.