Beyond Accuracy: Centroid-Guided Contrastive Loss for Structured Fraudulent Job Posting Detection

Syed Ali Ahmed; Malaika Raza; Muhammad Shoaib Siddiqui; Muhammad Rafi

cs.AISep 18, 2026

Beyond Accuracy: Centroid-Guided Contrastive Loss for Structured Fraudulent Job Posting Detection

Authors: Syed Ali Ahmed, Malaika Raza, Muhammad Shoaib Siddiqui, Muhammad Rafi

Organizations: National University of Computer and Emerging Sciences, Karachi, Pakistan · National University of Computer and Emerging Sciences, Karachi, 75020, Pakistan · Islamic University of Madinah, Madinah, Saudi Arabia · Faculty of Computer and Information Systems, Islamic University of Madinah, Madinah 42351, Saudi Arabia ·

Abstract

Fraudulent job posting detection aims to identify job advertisements that are corrupted either through fake content, misleading information, or negative intent, disrupting the online eco-system of job-seekers and employers. Existing studies in this domain lack effective methods to simultaneously achieve high accuracy and meaningful structure of latent-space representations that capture subtleties among fake posts. To this end, we propose Centroid-Guided Contrastive Loss (CGCL), a loss function which unifies classification with densely formulated clustering to consistently reshape latent-space through a centroid-driven top- $k$ push-and-pull mechanism. The complementary nature of CGCL enables the model to enforce accurate decision boundaries and maintain high clustering compactness, effectively capturing both class separability and latent structure. Extensive experiments demonstrate the state-of-the-art (SOTA) performance of our method on EMSCAD, a public benchmark dataset. The code associated with this work is available at: https://github.com/ali-ahmed925/CGCL_code

Figures & tables

Ref	Year	Feature Encoding	Imbalance Handling	Best Performing Model	Accuracy
[ 1 ]	2020	Categorical encoding	✗	Random Forest Classifier	98.27
[ 2 ]	2021	-	✗	BiLSTM	98.0
[ 3 ]	2021	TF-IDF	✗	MLP Classifier	71.0
[ 4 ]	2021	-	✗	Decision Forests	95.4
[ 5 ]	2022	TF-IDF	✓	Extra Tree Classifier	99.9
[ 6 ]	2022	BiLSTM	✗	BiLSTM	97.21

Table 1: Comparison of fake job postings detection methods showing feature extraction techniques, imbalance handling approaches, and performance metrics.

Column Name	Data Type	Non-Null Count	Description
job_id	int64	17880	Unique identifier for job
title	object	17880	Job position or role name
location	object	17534	City, state, or country posted
department	object	6333	Hiring department within company
salary_range	object	2868	Offered compensation range
company_profile	object	14572	Company background or description

Table 2: Job Dataset Schema: Column Specifications and Data Types

Panel of Figure 1 — Figure 1: Analysis of job postings: (a) Distribution of fraudulent vs. genuine postings, (b) Geographic distribution of postings, (c) Educational requirement analysis, and (d) Distribution by employment type

Figure 2: The pipeline consists of two phases: (1) Data Preprocessing & Augmentation, where raw text from the EMSCAD dataset is cleaned, vectorized using Word2Vec, and balanced using the ADASYN oversampling technique. (2) Deep Metric Learning, featuring a hierarchical Feed-Forward Neural Network (Input to FC-16) optimized by the proposed Centroid-based Geometric Contrastive Loss (CGCL). The CGCL formulation (bottom center) minimizes intra-class variance by pulling samples toward their respective centroids (C1, C2) while maximizing inter-class separability.

Panel of Figure 3 — Figure 3: Enhanced MLP Embedder Architecture. This diagram illustrates a seven-layer fully connected neural network used for feature extraction and latent representation learning.

Embedding	Epochs	Accuracy	Precision	Recall	F1-Score
Word2Vec	1500	0.981	0.97 / 0.98	0.98 / 0.98	0.98 / 0.98
	2000	0.985	0.98 / 0.99	0.99 / 0.99	0.98 / 0.99
	3000	0.986	0.98 / 0.99	0.99 / 0.99	0.98 / 0.99
GloVe	1500	0.985	0.98 / 0.99	0.99 / 0.99	0.98 / 0.98
	2000	0.981	0.97 / 0.98	0.98 / 0.98	0.98 / 0.98
	3000	0.990	0.98 / 0.99	0.99 / 0.99	0.98 / 0.99

Table 3: Performance of Word2Vec, GloVe, and TF-IDF embeddings under identical hyperparameters (

\textbf{

\alpha

}=0.6,\textbf{

\beta

}=0.9,k=3,m=1.0

). Values are reported as Macro / Micro . Best results are in bold .

Embedding	$\alpha$	$\beta$	$K$	Accuracy	F1
GloVe	0.5	0.7	5	0.992	0.987
	0.5	0.5	10	$0.987$	$0.978$
	0.5	0.5	7	$0.986$	$0.977$
	1.0	0.3	10	$0.986$	$0.976$
	1.0	0.5	7	$0.984$	$0.973$
Word2Vec	0.5	0.3	7	0.989	0.981

Table 4: Top hyperparameter configurations for GloVe and Word2Vec embeddings.

Embedding	$\boldsymbol{\alpha}$	$\boldsymbol{\beta}$	$\boldsymbol{k}$	Accuracy	F1-Score
Word2Vec	0.5	0.3	7	0.989	0.981
TF-IDF	0.6	0.9	3	0.991	0.990
GloVe	0.5	0.7	5	0.992	0.987

Table 5: Best performing models for each embedding method. Listed are the optimal hyperparameters and corresponding performance scores.

Metric	Score
Silhouette Score	0.701
Adjusted Rand Index (ARI)	0.953
Normalized Mutual Information (NMI)	0.903
Homogeneity Score	0.909
Completeness Score	0.898
V-Measure	0.903

Table 6: Clustering evaluation metrics for the final selected model (GloVe,

\alpha=0.5,\beta=0.7,k=5

Panel of Figure 4 — Figure 4: t-SNE (left) reveals the continuous internal structure of fraud templates, while UMAP (right) confirms the global separability and distinct sub-clustering of legitimate job categories.

Figure 5: Evolution of Latent Space. The progression from epoch 500 to 3000 shows the ’push’ mechanism creating a gradient of confidence for legitimate jobs (blue) while compacting fraudulent jobs (red) into a dense anomaly cluster. Hyperparameters: ( $\alpha=0.5,\;\beta=0.7,\;k=5$ )

Variant	Accuracy	Precision	Recall	F1	PR-AUC	MCC
CE_only	0.974 ±.002	0.763 ±.036	0.677 ±.015	0.718 ±.021	0.727 ±.022	0.706 ±.023
CE_weighted	0.976 ±.003	0.772 ±.043	0.719 ±.022	0.744 ±.026	0.740 ±.030	0.732 ±.028
CE_ADASYN	0.989 ±.000	0.967 ±.002	0.995 ±.003	0.981 ±.001	0.982 ±.003	0.973 ±.001
CE_CenterLoss	0.990 ±.001	0.968 ±.005	0.997 ±.002	0.982 ±.002	0.980 ±.006	0.975 ±.003
CGCL_full	0.983 ±.012	0.947 ±.035	0.997 ±.002	0.971 ±.020	0.981 ±.011	0.960 ±.028

Table 7: Test-set metrics (mean

\pm

std, five seeds, fraud class). Best per column in bold .

	Seed 7			Seed 13			Seed 21			Seed 42			Seed 99
Variant	F1	PR-AUC	MCC	F1	PR-AUC	MCC	F1	PR-AUC	MCC	F1	PR-AUC	MCC	F1	PR-AUC	MCC
CE_only	.727	.741	.716	.742	.749	.732	.689	.702	.675	.734	.744	.724	.696	.699	.681
CE_weighted	.751	.701	.741	.782	.780	.773	.732	.719	.718	.753	.770	.741	.702	.733	.688
CE_ADASYN	.980	.982	.972	.981	.982	.973	.980	.976	.972	.981	.986	.974	.982	.982	.974
CE_CenterLoss	.978	.975	.970	.980	.974	.972	.984	.981	.977	.983	.981	.977	.984	.991	.978
CGCL_full	.984	.990	.978	.980	.988	.972	.978	.984	.970	.981	.984	.974	.931	.961	.905

Table 8: Per-seed test-set results (F1 / PR-AUC / MCC). For each seed, best F1, best PR-AUC, and best MCC across variants are individually underlined .

Figure 6: Confusion matrices for all five variants (seed 7). Top row (L $\to$ R): CE_only, CE_weighted, CE_ADASYN. Bottom row: CE_CenterLoss, CGCL_full.

Contrastive Learning

Contrastive learning trains representations by bringing related examples together and separating unrelated ones for recognition, retrieval, and transfer.

See all 325 papers

Detecting Deceptive Recruitment: A Signal-theoretic Machine Learning Framework for Early Identification of Labour Exploitation
Sep 17, 2026Sajid Siraj, Mahnaz Hosseinzadeh, Amin Vafadarnikjoo +1
Leeds University Business School, University of Leeds, Leeds, UK · COMSATS University, Wah Campus, Islamabad, Pakistan · Sheffield University Management School, University of Sheffield, Sheffield, UK +1
Enhancing Online Recruitment with Category-Aware MoE and LLM-based Data Augmentation
Apr 23, 2026Minping Chen, Bing Xu, Zulong Chen +4Data Augmentation Mixture of Experts
HKUST (GZ) · Alibaba Group · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) +3
Layout-Aware Representation Learning for Open-Set ID Fraud Discovery
Apr 17, 2026Jinxing Li, Nicholas Ren, Cathy Chang +2Image Forgery Detection Visual Representation Learning
Persona Identities

Sep 17, 2026J/K move · Enter open · S save

Detecting Deceptive Recruitment: A Signal-theoretic Machine Learning Framework for Early Identification of Labour Exploitation

Sajid Siraj, Mahnaz Hosseinzadeh, Amin Vafadarnikjoo, Shuyang Li

Leeds University Business School, University of Leeds, Leeds, UK · COMSATS University, Wah Campus, Islamabad, Pakistan · Sheffield University Management School, University of Sheffield, Sheffield, UK+1

Open paper AskPDF arXiv

Deceptive online job advertisements have emerged as a primary pathway into forced labour, yet systematic detection methods remain underdeveloped due to data scarcity and absence of empirically validated indicators. We formalise this detection challenge as a classification problem under signalling theory, where exploiters transmit costless signals mimicking legitimate communications across textual, visual, and structural dimensions. Using 464 verified cases (164 deceptive, 300 legitimate) collected through anti-slavery charities across nine origin countries and 21 industries, we develop multimodal detection models combining computer vision, natural language processing, and semantic embeddings. Through systematic feature ablation experiments and repeated stratified cross-validation, we demonstrate that individual modalities achieve substantial discriminatory power (ROC-AUC: 0.87--0.97), whilst their integration yields modest further gains. SHAP-based analysis reveals that text quality and domain-specific risk language are the primary discriminators, with readability indices, risk keyword density, and visa sponsorship mentions ranking highest, followed by visual colour and texture features. These production quality gaps reflect resource constraints that prevent exploiters from maintaining professional standards across all communication channels simultaneously. We operationalise findings through a proof-of-concept decision support system providing interpretable risk scores for practitioners. This work demonstrates how rigorous analytical frameworks can address complex humanitarian operations challenges characterised by information asymmetry and limited ground-truth data.

Abstract

Contrastive Learning

Explore similar work

Detecting Deceptive Recruitment: A Signal-theoretic Machine Learning Framework for Early Identification of Labour Exploitation

Enhancing Online Recruitment with Category-Aware MoE and LLM-based Data Augmentation

Layout-Aware Representation Learning for Open-Set ID Fraud Discovery