Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.22907887# PI-DGAT — Policy-Invariant Dynamic Graph Attention Network Companion code and dataset for the study**PI-DGAT: A Policy-Invariant Dynamic Graph Attention Network for DynamicEarly Warning of Credit Default Risk of Small and Micro Enterprises**. PI-DGAT is a dynamic multi-relational graph model with a dual-channelattention operator that decouples an invariant channel from anenvironment channel, an interventional environment-reweighting trainingobjective enforcing invariance across policy cycles, a censoring-awarediscrete-time hazard rate head, and a non-exchangeable conformaldecision layer that outputs an admit / review / reject decision with astatistically controlled false-rejection rate. This repository releases the complete implementation together with thedataset used to produce the reported tables and figures. --- ## Repository contents Two archives are attached to the GitHub release. They are meant to bedownloaded together into the same working directory. | Archive | Contents ||-----------------------------|-------------------------------------------------------------------------------------------------------|| **`PI-DGAT_dataset.zip`** | Eighteen CSV files, one data-dictionary README. 218,436 firm nodes, 1,912,364 firm-quarter observations, 787,522 valid early-warning samples across four policy environments (E1–E4). || **`PI-DGAT_code.zip`** | PyTorch implementation of every equation (1)–(21) and every algorithm (1–3); baselines (LightGBM, GAT, HAN, EvolveGCN, DySAT, DyMGNN); driver scripts for Tables 5–7, the ablation study, the OOD-generalisation study, the conformal-coverage study, the economic-cost study, and Figures 8–13 (Nature style). | The two archives together reproduce every reported quantitativeresult. --- ## Quick start ```bash# 1. Download and unpack both archives side by sideunzip PI-DGAT_dataset.zip # → PI-DGAT_dataset/unzip PI-DGAT_code.zip # → PI-DGAT_code/ # 2. Install dependenciescd PI-DGAT_codepython3 -m venv .venvsource .venv/bin/activatepip install -r requirements.txt # 3. One command runs every experimentbash run_all.sh``` `run_all.sh` executes eight steps and writes every output under`PI-DGAT_code/results/`: 1. Pipeline verification — end-to-end run on the shipped dataset that exercises every module and confirms loss descent, finite gradients and hazard rates strictly inside (0, 1).2. Table 5 — early-warning performance comparison on test set E4.3. Table 6 — out-of-distribution generalisation E1 / E2 / E3 → E4.4. Table 7 — nominal versus empirical false-rejection rates for the non-exchangeable conformal decision layer.5. Ablation table — dual-channel decoupling, interventional reweighting, self-preserving residual, and conformal decision layer.6. Supporting tables for the attention-shift, lead-time and PCA analyses.7. Cost-of-misjudgment analysis reproducing the 18.8 % reduction in aggregate cost relative to the strongest baseline.8. Figures 8–13 rendered in Nature style (white background, English, ≥ 20 pt smallest text). --- ## Method overview Given a dynamic multi-relational credit graph with four edge types(guarantee, transaction, equity / executive, regional-industryproximity) and a firm-level environment vector composed of policytiming, pilot-zone affiliation, industry environmental standardstringency and firm-level green exposure, PI-DGAT 1. **Decouples** the invariant channel and the environment channel inside the scoring function of the graph attention operator, so information about repayment capacity is not mixed with information sensitive to policy (Equations (5)–(10)).2. **Enforces invariance** by intervening on the environment channel at training time and adding a cross-environment variance and gradient- norm penalty (Equations (11) and (15)).3. **Estimates a hazard-rate sequence** through a discrete-time, censoring-aware head with a per-period baseline (Equations (12)–(14)).4. **Converts hazard rates into decisions** through a non-exchangeable conformal quantile with time-decay weighting and a buffer band whose width is aligned to review-capacity constraints (Equations (18)–(21)). Full-scale training uses the following configuration: - 4 096 center nodes per mini-batch, subgraph fan-outs (25, 10);- 120 epochs, Adam, learning rate 1e-3;- λ₁ = 0.5 (invariance term), λ₂ = 0.1 (decorrelation term);- α = 0.10 (nominal false-rejection tolerance);- ζ = 0.85 (time-decay factor in the weighted conformal quantile);- five independent training runs, means and standard deviations reported. A single full run takes about 3.7 hours on one NVIDIA A100 40GB, with apeak memory footprint of about 31 GB. --- ## Reported results Headline numbers from the reported tables (see`PI-DGAT_code/results/tables/`): - **Table 5 (test set E4, mean ± std over five runs).** PI-DGAT reaches AUC 0.864 ± 0.007, KS 0.552 ± 0.009, AP 0.283 ± 0.008 and an early-warning lead time of 2.34 ± 0.19 quarters, best on all four metrics.- **Table 6 (out-of-distribution AUC).** Training on E1 alone and testing on E4, PI-DGAT reaches AUC 0.831 versus 0.762 for the strongest baseline; the decay ratio is roughly 2.4× smaller.- **Table 7 (conformal coverage).** At the nominal level α = 0.10, PI-DGAT's empirical false-rejection rate is 0.108 with a standard deviation of 0.006, versus 0.147 under the standard conformal procedure.- **Cost analysis.** Under matched operating points, PI-DGAT reduces the aggregate cost of misjudgment by 18.8 % relative to the strongest baseline, of which 2.5 pp comes from lower missed-default losses and 16.3 pp from lower opportunity cost through tighter false-rejection control. --- ## Data provenance The dataset is assembled from five public streams: firm registry andbusiness-change records from `gsxt.gov.cn`, guarantee and subrogationevents from the Credit Reference Center of the People's Bank of China,upstream–downstream trade linkages from customs and stock-exchangedisclosures, industry environmental-standard stringency from NBS andMEE releases, and green-finance policy and pilot-zone information fromPBoC and State Council designations of the Green Finance Reform andInnovation Pilot Zones (three batches: 2017Q3, 2019Q4, 2022Q3). Sectoral encoding follows GB/T 4754-2017. Administrative-divisionencoding follows GB/T 2260. Small-and-micro sizing follows the NBSStatistical Method for Classifying Large, Medium, Small and MicroEnterprises (2017). Personal identifiers are not stored anywhere in thearchive. --- ## Citation If this work is useful to your research, please cite the study. ```bibtex@article{pidgat2026, title = {A Policy-Invariant Dynamic Graph Attention Network for Dynamic Early Warning of Credit Default Risk of Small and Micro Enterprises}, author = {\{corresponding author\}}, year = {2026}}``` --- ## License Released under the MIT License. Third-party dependencies retain theirown licenses.
Keyword(s):
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)