Fraudulent job posting detection aims to identify advertisements that contain fake content, misleading information, or malicious intent, which disrupts the online ecosystem for job seekers and employers. Existing approaches typically focus on classification accuracy while neglecting the latent‑space structure needed to capture subtle differences among fake posts. To address this gap we introduce Centroid-Guided Contrastive Loss (CGCL), a loss that unifies classification and dense clustering. CGCL continuously reshapes the latent space via a centroid‑driven top‑$k$ push‑and‑pull mechanism:
- Compute a feature centroid for each class,
- For each sample, select the $k$ nearest negative centroids and push the sample away,
- Pull the sample toward its positive centroid,
- Combine these contrastive terms with the standard cross‑entropy loss.
This formulation enforces accurate decision boundaries while preserving high clustering compactness, thereby achieving both class separability and meaningful latent structure. Extensive experiments on the public benchmark EMSCAD demonstrate that CGCL attains state‑of‑the‑art performance across accuracy, F1, and clustering metrics such as Silhouette score. The implementation is released at https://github.com/ali-ahmed925/CGCL_code for reproducibility and further exploration.
Review