Imputation of missing values of a time series based on joint application of analytical algorithms and neural networks

Currently, time series data are processed in a wide range of scientific and practical applications, where the imputation of points or blocks missing due to hardware/software failures or the human factor is topical. In the article, we present the SANNI (Snippet and Artificial Neural Network-based Imputation) method to recover the missing values of the time series processed offline. SANNI includes two neural network models, namely Recognizer and Reconstructor. The Recognizer determines the snippet (typical subsequence) of the time series that a given subsequence with a missing point is the most similar to. The Recognizer consists of the three groups of layers: convolutional, recurrent, and fully connected. The Reconstructor, using the Recognizer's output and a subsequence with a missing point, restores the missing point. The Reconstructor consists of three groups of layers: convolutional, recurrent, and fully connected. The topology of the Recognizer and Reconstructor layers is parameterized with respect to the snippet length. We also present a way to prepare training sets for the Recognizer and Reconstructor. Our computational experiments showed that among the state-of-the-art analytical and neural network imputation methods, SANNI is among the top three.