Sensitivity of CGCL training to graph topology and contrastive temperature

Investigate how the interaction between graph topology and the contrastive temperature hyperparameter affects convergence speed and training stability in the Centroid-Guided Contrastive Loss (CGCL) model for fraudulent job-posting detection.

Background

The paper evaluates the full Centroid-Guided Contrastive Loss (CGCL) model against several ablation variants for fraudulent job-posting detection. Although CGCL achieves strong recall and PR-AUC, its performance exhibits greater variation across random seeds than the cross-entropy-based alternatives.

In particular, the results indicate that one seed converged substantially more slowly. The authors associate this behavior with a possible interaction between graph topology and the contrastive temperature hyperparameter, but do not determine the mechanism or characterize the conditions under which this sensitivity occurs. They explicitly identify this issue as an open problem for future work.

References

Seed~99 converged significantly more slowly, suggesting sensitivity to the interaction between graph topology and the contrastive temperature hyper-parameter---an open problem for future work.

— Beyond Accuracy: Centroid-Guided Contrastive Loss for Structured Fraudulent Job Posting Detection  (2609.21599 - Ahmed et al., 18 Sep 2026) in Section “Ablation Study,” subsection “Discussion,” paragraph “CGCL graph context and variance”