Residual SemGCN-MDN: A Hybrid Graph Neural Network Approach for Time Series Forecasting

Abstract

Time series forecasting on graph-structured data is an important problem in various fields such as tracing diseases, teaching, and social networks. Graph Neural Network (GNN)-based models, presently used in most applications, often assume a unimodal distribution for future predictions, making it incapable of grappling with the uncertainty and multi-modal nature of real-life datasets. In this paper, we propose a novel residual learning-based hybrid technique that integrates the Semantic Graph Convolutional Network (SemGCN) with the Mixture Density Network (MDN) output layer. The method breaks down the forecasting process into two parts: (i) the linear regression model predicts the deterministic trend, and (ii) SemGCN & MDN probabilistically models the residual part. This framework permits the model to recognize the graph’s structural dependencies as well as the uncertainty in the temporal dynamics. The combined method is evaluated using several benchmark datasets of graph-based time series and shown to attain competitive performance in terms of Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Scaled Error (MASE) while providing a distribution-aware prediction mechanism for uncertainty quantification.