A Novel Discrete Time Series Representation With De Bruijn Graphs for Enhanced Forecasting Using TimesNet

In this paper, we present a novel method for advancing time series forecasting by representing discretized time series data through de Bruijn Graphs (dBGs). This method harnesses the capability of dBGs to encapsulate and project future states from historical sequences, thus enhancing predictive analytics in time series. Our approach is multi-faceted, involving: (1) encoding time series data as a dBG; (2) the application of graph representation learning, specifically <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">struct2vec</i>, to distill salient features from dBG constructed from time series and (3) the seamless integration of these extracted features into the state of the art TimesNet model to bolster short-term forecasting accuracy. Empirical evaluations conducted on the M4 datasets illustrate that our approach not only maintains the intrinsic dynamics of the time series but also achieves notable improvements in forecasting performance across diverse datasets. All the code developed for this study can be found at: https://github.com/KurbanIntelligenceLab/dBGTime-Series-Library.