Multi-Domain Counterfactual Causal Graphs for Spurious Pathway Detection and Functional Risk Estimation

Deep neural networks (DNNs) are prone to exploiting spurious correlations, especially when trained on multi-source datasets, where pseudo-causal paths can form across domains and interfere with generalization. This work introduces a method for detecting such paths and quantifying their influence through a counterfactual causal graph framework. By assembling cross-domain causal graphs from datasets like DomainNet (six domains, 18,500 samples) and OfficeHome (four domains, 17,000 samples), we incorporate both statistical correlations (based on Pearson coefficients) and expert-defined priors. A causal reasoning model is constructed on top of a ResNet-18 backbone, trained with a learning rate of 0.0008 and batch size of 64. After 40 epochs, the model achieves an AUC of 0.92 in distinguishing between true and pseudo-causal signals. To quantify shortcut interference, we propose the Functional Shortcut Risk Index (FSRI), which combines path scoring and intervention-based accuracy gain, weighted at 0.65 and 0.35, respectively. Using FSRI to guide path weight adjustment improves Top-1 transfer consistency in target domains from 65.2% to 83.5%. On average, accuracy on DomainNet and OfficeHome increases by 13.8% and 15.6%, with statistical significance confirmed by paired t-tests (p < 0.001). Further analysis of layer-wise activations shows that when true causal paths are used, activation in the fourth residual block and final classifier reaches 0.82 and 0.88, while pseudo-paths yield much lower values (0.28 and 0.36). These results highlight the potential of causal graph diagnostics in mitigating shortcut learning and improving model robustness across domains.