Neural networks generative models for time series

Nowadays, time series are a widely-exploited methodology to describe phenomena belonging to different fields. In fact, electrical consumption can be explained, from a data analysis perspective, with a time ser-ies, as for healthcare, financial index, air pollution or parking occupancy rate. Applying time series to dif-ferent areas of interest has contributed to the exponential rise in interest by both practitioners and academics. On the other side, especially regarding static data, a new trend is acquiring even more rele-vance in the data analysis community, namely neural network generative approaches. Generative approaches aim to generate new, fake samples given a dataset of real data by implicitly learning the prob-ability distribution underlining data. In this way, several tasks can be addressed, such as data augmenta-tion, class imbalance, anomaly detection or privacy. However, even if this topic is relatively well -established in the literature related to static data regarding time series, the debate is still open. This paper contributes to this debate by comparing four neural network-based generative approaches for time series belonging to the state-of-the-art methodologies in literature. The comparison has been carried out on five public and private datasets and on different time granularities, with a total number of 13 experimental scenario. Our work aims to provide a wide overview of the performances of the compared methodologies when working in different conditions like seasonality, strong autoregressive components and long or short sequences.(c) 2022 The Authors. Published by Elsevier B.V. on behalf of King Saud University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Automated summary

Time series is a widely-used methodology to describe phenomena in various fields. Neural network generative approaches, which aim to generate new samples based on real data by learning the underlying probability distribution, are gaining more relevance in the data analysis community. This paper contributes to the debate by comparing four neural network-based generative approaches for time series, evaluating their performances under different experimental conditions.