Replication Data for "Stable Predictions, Fragile Coordination: Diagnosing Interpretation Risk in AI Decision Support Systems"

Description

This dataset contains the analytical data and replication code for the DIV-DRO framework study on interpretation stability in AI-assisted decision support systems, applied to 544 forest fire incidents in Taiwan (FANCA, 2024). Includes cleaned event data, 20-seed split experiment results, pairwise Coverage Loss and Narrative Shift metrics (190 pairs), and distributionally robust optimization outputs.

Steps to reproduce

1. Run DIV_DRO_complete_v4_with_supplementary.py to reproduce all DIV and DRO analyses. 2. Input data: historical_fire_dataset_V2_clean_en.csv (544 forest fire incidents, Taiwan FANCA 2024). 3. Set model_seed=42 for shared test set (n=136). Split seeds: 1,3,7,12,17,21,28,35,42,50,57,63,70,77,84,91,93,95,99,100. 4. Output tables match Table_DIV_Summary_20seeds.csv, Table_WC_CL_Full/Proxy_20seeds.csv, and Table_Threshold_Full/Proxy_20seeds.csv. 5. Run Plot_figures.py to reproduce Figures 5-7.

Institutions

• National Pingtung University of Science and Technology

  Taiwan, Pingtung City

Categories

Licence