From an AI/network science perspective, there are several serious methodological pitfalls with the concept of ‘relational risk’.
Criticism: Causality confusion and measurement imperfection.
The most fundamental problem is the causal structure of the claim that “signals of [[executive network]] dissolution are caught in advance.” Converting correlation to causality requires controlling for the possibility of reverse causality. Unless the Granger causality test or instrumental variable (IV) estimation is used to verify whether the financial crisis causes network disintegration or the other way around, this figure is merely a moving indicator. The figure of 85.9% itself is suspicious of sample selection bias (survivorship bias). Analyzing only companies with trading suspensions excludes companies that survived even if there was a network dissolution.
There is also the issue of [[network centrality]] metric selection. Measuring hub convergence only through simple degree centrality or betweenness centrality does not capture temporal changes in dynamic networks. The concurrent executive network must be analyzed as a temporal graph rather than a static snapshot to capture the speed and direction of structural change before the crisis.
Improvement proposal: AI-based [[Relational Risk]] measurement pipeline
We propose the following three-step structure:
Step 1 — Build a dynamic knowledge graph: Integrate Financial Supervisory Service disclosure data, corporate registration information, and shareholder list to construct a heterogeneous temporal graph with a time axis. Nodes are defined as executives, companies, and financial institutions, and edges are defined as concurrent positions, shares, and loans.
Step 2 — GNN-based risk score calculation: Apply Temporal Graph Neural Network (T-GNN) to extract the temporal embedding of each node and learn the network structure collapse pattern. Key design factors are edge extinction rate, cluster coefficient reduction rate, and intermediary hub departure pattern.
Step 3 — Ensuring interpretability through
This structure converts the macro-structural proposition r > g into a form that can be verified with micro-network data. For relational risk to function as a policy tool, this level of measurement infrastructure must be in place.
supporting data
- Kipf, T. N., & Welling, M. (2017). Semi-supervised classification with graph convolutional networks. ICLR.
- Xu, D. et al. (2020). Inductive representation learning on temporal graphs. ICLR.
- Granger, C. W. J. (1969). Investigating causal relations by econometric models. Econometrica, 37(3), 424–438.
- Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. NeurIPS.
- Heckman, J. J. (1979). Sample selection bias as a specification error. Econometrica, 47(1), 153–161.
- Borgatti, S. P., & Everett, M. G. (2006). A graph-theoretic perspective on centrality. Social Networks, 28(4), 466–484.
#relationalrisk #raymondsrisk #raymondsindex #konnectai
Comments
Post a Comment