GeneaLink: Reconstructing Marriages in Genealogies with Visual Analytics Coupling GNN Link Prediction and Multi-Agent Negotiation
Authors
Zhuoyang Bu (ShanghaiTech University), Junjie Xiong (ShanghaiTech University), Shenghan Gao (ShanghaiTech University), Haipeng Zhang (ShanghaiTech University), Quan Li (ShanghaiTech University)
Presentation
- Session
- Tools of the Trade
- Time
- Thursday, Nov 12, 13:45 – 13:54 (US/Eastern) · session 13:00 – 14:30
- Location
- Hall Essex center
Keywords
visual analytics, genealogy reconstruction, heterogeneous graph learning, LLM multi-agent systems
Abstract
In historical studies, the recording convention of patrilineal genealogies systematically conceals the cross-clan connections formed by women through marriage, resulting in the structural disintegration of ancient kinship networks into isolated family trees. Existing graph neural network link prediction methods can generate candidate marriage pairs, but they cannot provide interpretable reasons; and relying solely on large language model reasoning is difficult to process such a large-scale and highly complex data structure in a single round, leading experts to find it difficult to trust or review recommended results. We introduce GeneaLink, a complementary and reversible interactive visual analysis system that combines Heterogeneous Graph Transformer (HGT) and multi-round Multi-agent Simulation, enabling historians to examine model confidence, trace the source of predictions, inject domain knowledge into the deliberation process of agents, and confirm or revoke decisions under complete audit trails. An expert-guided system walkthrough demonstrates the effectiveness of GeneaLink.
For Practitioners
Target Audience: This paper will primarily interest visual analytics researchers, digital humanities scholars (particularly computational historians and demographers), and data scientists working at the intersection of graph neural networks and large language model agents. Application in Practice: For VA designers: Can adapt GeneaLink’s coupling of graph learning with multi-agent simulations, particularly the "prior transfer" mechanism, to build transparent human-in-the-loop interfaces for auditing algorithmic decisions. For computational historians & demographers: Can adopt this MAS-driven methodology to address systemic archival biases—such as missing female marital ties—by modeling historical marriage formation as a testable negotiation process rather than a static link prediction problem. For data scientists & AI engineers: Can leverage this system's architecture, which grounds LLM reasoning in deterministic structural evidence, to mitigate unfaithful rationalizations and build steerable multi-agent systems for complex domain-specific tasks.