Learning Visibility Attention Graph Representation for Time Series Forecasting

Visibility algorithm acts as a mapping that bridges graph representation learning with time series analysis, which has been broadly investigated for forecasting tasks. However, the intrinsic nature of visibility encoding yields graphs structured exclusively by binary adjacency matrix, leading to inevitable information loss of temporal sequence during the mapping. To this end, we introduce Angular Visibility Graph Networks (AVGNets), designed with two core features: (i) The framework reconstructs weighted graphs to encode time series by leveraging topological insights derived from visual angles of visibility networks, which capture sequential and structural information within weighted angular matrix. (ii) ProbAttention module is proposed for evaluating probabilistic attention of weighted networks, with remarkable capabilities to extract intrinsic and extrinsic temporal dependencies across multi-layer graphs. Extensive experiments and ablation studies on real-world datasets covering diverse ranges demonstrate that AVGNets achieve state-of-the-art performance, offering an innovative perspective on graph representation for sequence modeling.