A meta-learning based neural network and LSTM for univariate time series missing data imputation

Time series are regularly collected data that describe the average evolution of an event over time, making them increasingly relevant in areas such as business, natural sciences and medicine. A major challenge related to time series is data loss, and several approaches have been developed to recover missing values in univariate time series (UTS). This work aims to improve the imputation of missing data in univariate and heterogeneous time series. Thus, we built a diverse database covering different time series domains and selected a set of data imputation techniques. The results show that imputation in time series is challenging, especially due to the variability of the series, the position of missing data and the number of samples passed to each technique. The HybridLSTM network, developed in this study, proved effective in recommending the most suitable imputation techniques for each series, resulting in a lower average error than using a single technique or recent techniques such as Pix2Pix and Moment. In addition, adopting a hybrid loss function, which considers multi-class and multi-label tasks, contributed to optimal or near-optimal performance, even in cases where the ideal was not achieved. These advances were possible thanks to the efficient but simple construction of metadata and the innovative approach of locally combining several imputation techniques within the same series. We observed that meta-learning has great potential to be applied in real contexts where the ideal technique is not previously known and the data has not been pre-treated in terms of data values. Moreover, as our experiments were very close to this context, it became useful, as the model performed very close to the ideal, validating the applicability of the adaptive meta-learning approach to optimize the imputation of missing data in real contexts.