Long Short-term Cognitive Networks: An Empirical Performance Study

Long Short-term Cognitive Networks (LSTCNs) are recurrent neural networks for univariate and multivariate time series forecasting. This interpretable neural system is rooted in cognitive mapping formalism in the sense that both neural concepts and weights have a precise meaning for the problem being modeled. However, its weights are not constrained to any specific interval, therefore conferring to the model improved approximation capabilities. Originally designed for handling very long time series, the model's performance remains unexplored when it comes to shorter time series that often describe realworld applications. In this paper, we conduct an empirical study to assess both the efficacy and efficiency of the LSTCN model using 25 time series datasets and different prediction horizons. The numerical simulations have concluded that after performing hyper-parameter tuning, LSTCNs are as powerful as state-ofthe-art deep learning algorithms, such as the Long Short-term Memory and the Gated Recurrent Unit, in terms of forecasting error. However, in terms of training time, the LSTCN model largely outperforms the remaining recurrent neural networks, thus emerging as the winner in our study.