Decoupling time and space: An adaptive shared graph convolutional network for dynamic market price forecasting

Accurate forecasting of product prices is vital for enterprises to anticipate market trends and adjust sales strategies proactively. Given the volatility of supply and demand, product prices often exhibit complex spatiotemporal coupling characteristics. However, traditional deep learning methods like graph neural networks mainly focus on local feature learning, limiting their ability to model the continuous evolution of market dynamics and spatial diffusion patterns. To overcome this limitation, this study proposes a novel spatiotemporal decoupling adaptive-shared graph convolutional network (STDAsh-GCN) for product price prediction. First, STDAsh-GCN introduces a globally shared parameter mechanism in the graph convolution process, enabling deep decoupling of spatial and temporal representations. Furthermore, an adaptive feature aggregation module is designed to dynamically assess and enhance the contribution of each node during convolution, thereby improving the capacity of the model to capture and integrate salient features. To further reveal intrinsic structural dependencies, a shared attention mechanism is incorporated to balance the influence of input features and adjacency relationships. Extensive experiments on three real-world industrial datasets, including one from a potassium sulfate production enterprise, validate the effectiveness of the proposed method compared to existing state-of-the-art methods.