Imputation of multivariate time series based on the behavioral patterns and autoencoders

Currently, in a wide range of subject domains, the problem of imputation missing points or blocks of time series is topical. In the article, we present SAETI (Snippet-based Autoencoder for Time-series Imputation), a novel method for imputation of missing values in multidimensional time series that is based on the combined use of autoencoders and a time series of behavioral patterns (snippets). The imputation of a multidimensional subsequence is performed using the following two neural network models: The Recognizer, which receives a subsequence as input, where the gaps are pre-replaced with zeros, and determines the corresponding snippet for each dimension; and the Reconstructor, which takes as input a subsequence and a set of snippets received from the Recognizer, and replaces the missing elements with plausible synthetic values. The Reconstructor is implemented as a combination of the following two models: An Encoder that forms a hidden state for a set of input sequences and recognized snippets; and a Decoder that receives a hidden state as input, which imputes the original subsequence. In the article, we present a detailed description of the above models. The results of experiments over time series from real-world subject domains showed that SAETI is on average ahead of state-of-the-art analogs in terms of accuracy and shows better results when input time series reflect the activity of a certain subject.