Enterprise AI Analysis

The core methodology involves constructing heterogeneous hypergraphs to represent complex relationships (e.g., sector-based, fund-holding). Different HGN architectures are compared: HGTAN (tri-attention: intra-hyperedge, inter-hyperedge, inter-hypergraph), STHAN-SR (Hawkes-based temporal attention), HGTS-Former (patchification, hierarchical intra/inter-channel hypergraphs), and MSHyper (multi-scale hypergraph construction). Temporal encoding via attention captures historical patterns, while spatial modeling via hypergraphs captures group dependencies. Evaluation is based on classification metrics and financial performance.

HGTAN's Tri-Attention Network integrates relational information through three levels: intra-hyperedge (stock contributions to hyperedge), inter-hyperedge (hyperedge contributions to node), and inter-hypergraph (different hypergraph types). HGTS-Former uses patchification and token embedding, followed by multi-head self-attention for temporal representation, and then constructs a two-level hierarchical hypergraph for intra-channel and inter-channel relations. MSHyper employs multi-scale hypergraphs, constructing hyperedges based on anchor selection and top-k neighbors in feature space to capture dependencies across different temporal scales.

Future research will focus on developing approaches to incorporate dynamic hypergraph learning, allowing relationships among stocks to evolve over time. Additionally, integrating multi-modal signals such as financial news or sentiment embeddings into the models is a promising avenue to further enhance predictive capabilities and capture a more comprehensive market view.