VISIBILITY GRAPH NETWORKS FOR TIME SERIES FORECASTING

Abstract

The use of visibility graph networks (VGNs) for time series forecasting has emerged as an innovative paradigm, bridging traditional methods with modern graph-based approaches. From fundamental principles to cutting-edge methodologies, this thesis investigates the potential of VGNs across various aspects, each contributing to the development and refinement of forecasting techniques. The initial study introduces the foundational concept of VGNs, employing visible nodes and node degree properties to map time series data into network structures. It establishes the groundwork for subsequent investigations by demonstrating the efficacy of these structures in forecasting scenarios, especially in limited data situations. Building upon this foundation, the research expands the VGN framework by emphasizing the importance of historical knowledge encoded within these networks. By introducing a novel similarity index, this work captures past patterns and current data movements, enhancing predictive capabilities significantly beyond existing methods. Advancing further, we extend the VGN approach and propose MultiNets, a multivariate forecasting network that addresses information and correlation loss. This model introduces multiplex networks and intra-inter layer similarities, enabling a deeper analysis of complex temporal dependencies across variables. The superior performance in carbon price forecasting demonstrates the potential of graph-based representations. Finally, we present angular visibility graph, addressing information loss by reconstructing weighted graphs using angular matrix-based encoding. This innovative framework introduces ProbAttention module for evaluating probabilistic attention distributions across multi-layer graphs, excelling particularly in long-term forecasting scenarios. Collectively, this sequential progression of studies within the thesis underscores a seamless evolution from foundational concepts to pioneering methodologies within the domain of VGNs for time series forecasting. Beyond enhancing predictive accuracy, these findings set a new trajectory for future research in this dynamic field, offering a promising avenue for further exploration and refinement.

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Xiaojun Zeng (Supervisor) & Ke Chen (Supervisor)