GAR-HGNN: A Heterogeneous Graph-Based Approach to Recruitment Fraud Detection

In this paper, we propose a gated attention and dual feature enhancement method GAR-HGNN based on heterogeneous graph to address the problem of low recall and F1 value in recruitment fraud detection. By dynamically adjusting the fusion weights of text and attribute features through the gated attention mechanism, combining the labour market and criminology theories, we design industry-salary anomaly index (node attribute) and geography-industry risk density (edge weight) to enhance the graph structure risk expression, and optimize the training of imbalanced data by using the weighted loss function based on the anomaly index. Experiments show that the method achieves an accuracy of 0.9896, precision of 0.9626, recall of 0.9962, and F1 value of 0.9791 on the EMSCD dataset.