BalancedCL: A Balanced Contrastive Learning Framework for Process Mining with Multi-Head Attention

Process mining has become a key tool to extract insights from event logs in business processes. However, traditional clustering techniques struggle with imbalanced data and capturing key decision points. To address these limitations we propose BalancedCL, a new contrastive learning framework that combines multi-head attention and targeted data augmentation. The BalancedCL framework balances the representation of minority paths through balanced augmentation and focuses on influential events with multi-head attention. Additionally a weighted contrastive loss function is used to prevent frequent patterns to dominate, so all paths contribute equally. Experiments on BPI 2012 show that BalancedCL achieves a Silhouette Score of 0.6063, outperforming traditional clustering methods and deep learning approaches. Attention weights analysis reveals that decision and termination events are the key to differentiate process paths. This shows that BalancedCL is a robust and interpretable approach for process mining that can find meaningful patterns even with imbalanced data.