Reproduce “Multi-Graph Tensor Networks” on OpenResearch

About this paper

Yao Lei Xu, Kriton Konstantinidis, Danilo P. Mandic

Researchers at Imperial College London developed Multi-Graph Tensor Networks (MGTN), a framework integrating graph signal processing with tensor networks to efficiently process multi-modal data across multiple graph domains. Applied to FOREX algorithmic trading, the fMGTN variant achieved a 0.8018% total return with only 531 trainable parameters, outperforming GRU, TTNN, and GCN baselines while exhibiting better risk-adjusted returns and reduced model complexity.

Key insights

• The integration of graph signal processing (GSP) and tensor networks (TNs) allows MGTN to leverage graph structures for relational information and tensor decompositions for efficient high-dimensional data compression.
• Multi-linear graph filters can be generalized beyond time-series to capture complex interactions across various modalities represented as distinct graph domains.
• Tensor-Train decomposition effectively mitigates the 'Curse of Dimensionality' by reducing the parameter count of dense layers while preserving the multi-modal structure, crucial for handling noisy data like financial time series.

Summary by alphaXiv.