Next Event Prediction in Business Process Logs Using Stacked Autoencoders with N-gram Encoding and Feature Hashing

Abstract

Proactive monitoring of business processes has become a key competitive advantage for firms, enabling timely interventions to prevent workflow deviations. Process-aware information systems generate extensive logs, which serve as valuable resources for predictive analytics. In this context, this study presents a deep learning-based approach for predicting the next event in an ongoing process by analyzing historical execution logs. The proposed method formulates event prediction as a classification task, leveraging n-gram encoding and feature hashing for effective feature preprocessing. The model consists of a multi-stage deep learning framework, incorporating stacked autoencoders for unsupervised pretraining, followed by a supervised fine-tuning phase to optimize classification accuracy. Experimental validation was conducted on six real-life event log datasets, including BPI Challenge 2012 (subsets A, O, W), BPI Challenge 2013 (Incidents, Problems), and Helpdesk logs. The proposed approach achieved up to 83.1% accuracy, 85.2% precision, and 92.3% AUC on the BPI 2012_A dataset, outperforming stateof-the-art classifiers such as LSTM-based models and Bayesian regularized PFAs. Notably, it demonstrated a 6–11% improvement in recall over existing methods on key datasets. The results highlight the model’s ability to capture complex process dynamics and improve proactive situation awareness. Additionally, the study explores the impact of hyperparameter tuning and addresses data imbalance challenges using RBF-based synthetic data generation, contributing a robust framework for real-time decision support in business process management.