An Attention-Aware Graph Neural Network and Multi-Modal Transformer Architecture for Cross-Platform Multi-Source Data Fusion in E-Commerce Demand Forecasting

• Ada John

  ladoke Akintola university of technology

  Author

E-commerce demand forecasting increasingly requires the integration of heterogeneous data sources spanning transactional records, social media signals, visual product attributes, and cross-platform behavioral traces. Existing approaches predominantly rely on unimodal time-series modeling or shallow feature concatenation, failing to capture the complex relational structures and cross-modal dependencies that characterize modern digital commerce ecosystems. This study proposes an Attention-Aware Graph Neural Network and Multi-Modal Transformer (AAGNN-MMT) architecture designed to address these limitations through unified representation learning across heterogeneous modalities. The framework constructs a dynamic heterogeneous graph encoding inter-product, inter-category, and cross-platform relationships, while a multi-modal transformer applies cross-attention mechanisms to fuse textual, visual, and temporal representations. Evaluation on a multi-source e-commerce dataset comprising 2.3 million transaction records and 890,000 multimodal product interactions demonstrates that the AAGNN-MMT achieves 89.4% directional forecast accuracy, outperforming the strongest baseline by 12.7 percentage points in WAPE and reducing MAPE to 6.83%. Ablation studies confirm that attention-aware graph propagation contributes 34% of the performance gain, while cross-modal fusion accounts for 28%. The findings establish that relational and multimodal information integration is not merely additive but produces emergent predictive capabilities unavailable to single-source architectures. The study provides a replicable framework for practitioners seeking to deploy multi-source demand forecasting systems and advances theoretical understanding of attention mechanisms in heterogeneous data fusion contexts.

• pdf

Copyright (c) 2026 Ada John (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.