A DEEP LEARNING-BASED APPROACH TO ONLINE RECRUITMENT FRAUD DETECTION

Abstract

These days, the majority of businesses use digital platforms to seek new hires in order to streamline the recruiting process. Fraudulent advertising is a consequence of the sharp rise in the usage of online job posting platforms. The fraudsters use phoney job advertisements to get revenue. Fraud in online hiring has become a significant problem in cybercrime. Therefore, to eliminate online job frauds, it is essential to identify phoney job ads. The goal of this research is to employ two transformer-based deep learning models, namely Bidirectional Encoder Representations from Transformers and Robustly Optimised BERT-Pretraining Approach (RoBERTa), to accurately detect fake job postings. Traditional machine learning and deep learning algorithms have been used in recent studies to detect fake job postings. By combining job ads from three distinct sources, a unique dataset of fraudulent job advertisements is suggested in this study. The effectiveness of current algorithms to identify fake jobs is hampered by the outdated and restricted benchmark datasets, which are based on knowledge of particular job advertisements. We thus update it with the most recent job openings. The class imbalance issue in identifying phoney employment is brought to light by exploratory data analysis (EDA), which causes the model to behave aggressively against the minority class. The work at hand employs 10 of the best Synthetic Minority Oversampling Technique (SMOTE) variations in order to address this issue.
Analysis and comparison are done between the models' performances balanced by each SMOTE version. Every strategy that is used is carried out in a competitive manner. However, with a balanced accuracy and recall of almost 90%, BERT+SMOBD SMOTE produced the best results.