Collusion detection is a critical process for identifying instances in which entities, such as companies or individuals, secretly collaborate to engage in fraudulent practices, often characterized by agreements on prices, market shares, or tactics to avoid competition. GNNs explicitly allow the incorporation of network structures, which are prevalent in collusion and key to improving the detection rate. This talk tackles the challenge of detecting fraud patterns using GNNs and highlights how these models can effectively learn network structures inherent in bidding markets. We propose a fraud detection algorithm that uses relational graph convolutional networks (R-GCNs) to analyze the inherent network structures in bidding data. R-GCNs have the ability to assign different weights to different types of edges. By using this extension of GNNs, different types of relationships between the bids can be combined to generate richer node embeddings, making them more effective at detecting collusive behavior among companies participating in bids and tenders. We develop and train these models using the PyTorch framework and the Deep Graph Library (DGL), applying them to datasets from Japan, the United States, Switzerland, Italy, and Brazil. Our empirical findings show that GNNs outperform traditional NNs in identifying complex collusive patterns, offering a new solution to this fraud problem. For data scientists, especially those working with network data, the detection of anomalous patterns in network structures is a critical challenge across different domains, e.g. financial fraud detection or social network analysis. Attendees will gain practical strategies for applying GNNs to classification tasks to detect patterns in network data and a deeper understanding of how graph embeddings can enhance fraud detection models.