پيش‌بيني سري زماني چندمتغيره با استفاده از شبكه عصبي گرافي با تمركز بر ساخت خودكار گراف از داده‌ها

Multivariate time series forecasting holds a special position in various research an‎d applied domains due to its ability to model complex data an‎d interdependencies among variables. This type of forecasting plays a crucial role in areas such as deman‎d management, resource allocation, an‎d financial market analysis. Among these, graph-based methods for multivariate time series forecasting exhibit high capability in modeling complex an‎d nonlinear dependencies among variables, which are abundantly found in real-world applications. However, modeling multivariate time series with high volatility poses specific challenges due to inherent complexities an‎d intertwined relationships. Such data are often observed in economic domains, including financial markets an‎d commodity prices. Graph neural networks have been introduced as a novel approach for modeling spatio-temporal dependencies in multivariate time series, but these methods typically face challenges such as heavy computational load an‎d high processing resource requirements. In this research, two proposed architectures titled SGMTSP an‎d AGMTSP are presented. The first architecture is designed based on explicit graph construction through measuring similarity among time series, while the second architecture models dependencies based on implicit graph construction utilizing an attention mechanism. In this research, efforts have been made to ensure that the proposed architectures can model multivariate time series with volatility at high accuracy, reduce computational complexity compared to existing models, an‎d provide an adaptive method for graph construction. The obtained results indicate that the proposed architectures have achieved significant improvements in performance eva‎luation metrics compared to comparative methods. In the Iranian stock market dataset, the SGMTSP architecture with 8% an‎d the AGMTSP architecture with 4% have shown improvements in the mean squared error metric compared to similar models.

5

148939