Predicting Process Behaviour using Deep Learning (1612.04600v2)

Abstract: Predicting business process behaviour is an important aspect of business process management. Motivated by research in natural language processing, this paper describes an application of deep learning with recurrent neural networks to the problem of predicting the next event in a business process. This is both a novel method in process prediction, which has largely relied on explicit process models, and also a novel application of deep learning methods. The approach is evaluated on two real datasets and our results surpass the state-of-the-art in prediction precision.

• The paper improves prediction precision by applying RNNs to forecast event sequences in business processes.
• It employs LSTM-based deep learning to implicitly model process behavior without traditional explicit methods.
• Experimental results on real datasets demonstrate over 80% accuracy, highlighting its potential for smart BPM.

An Overview of "Predicting Process Behaviour using Deep Learning"

The paper "Predicting Process Behaviour using Deep Learning," authored by Joerg Evermann, Jana-Rebecca Rehse, and Peter Fettke, presents a novel approach leveraging recurrent neural networks (RNNs) for predicting event sequences in business processes. The motivation behind this work stems from parallels between NLP and business process management (BPM), where predicting the next word in a sentence is akin to predicting the next event in a process. Traditional methods rely heavily on explicit models like state-transition diagrams, Hidden Markov Models (HMM), or Probabilistic Finite Automatons (PFA), which this research aims to refine and perhaps replace with implicit deep learning techniques.

Methodology

The authors examine the potential of deep learning through RNNs, specifically Long Short-Term Memory (LSTM) networks, to predict the next event in a process without relying on explicit process models. Events are contextualized as sequences, analogous to words in sentences, and entire logs of process events are thus treated as corpora. Their approach involves training RNNs on these event logs and evaluating the precision of model predictions. They assess the approach using two real-world datasets, one from a loan application process and another from IT incident management, highlighting the versatility and robustness of their method.

Key Contributions

The paper makes several important contributions to the field:

1. Improving Prediction Precision: RNN models reportedly achieve superior prediction precision compared to state-of-the-art methods, notably surpassing those based on probabilistic models like PFA.
2. Implicit Process Modelling: It illustrates the feasibility of using deep learning to predict process behavior without explicit models, which can be crucial for scenarios where these models are hard to derive or maintain.
3. Applicability in Smart BPM: The work supports the adoption of artificial intelligence techniques in business process management, pushing the boundaries of how intelligence and machine learning can be applied to solve complex prediction problems in BPM.

Experimental Evaluation

The experiments involve evaluating prediction precision on training and validation datasets, showing significant performance with precision often exceeding 80%. The research also examines the impact of incorporating additional information, such as resources, into event prediction — a feature that can enhance predictive prowess although potentially at the risk of overfitting without careful management.

Moreover, the authors explore different dimensions within the network's embedding space and varying the number of LSTM cells in the unrolled network to assess their impact on performance. The sensitivity analysis conducted reinforces the paper's assertion that an appropriately configured deep learning model, with mindful parameter settings, can achieve high predictive accuracy.

Implications and Future Work

This research opens avenues for a shift in how BPM might increasingly depend on learning from historical process behaviors to anticipate future actions, thereby bolstering operational efficiency and decision-making processes. From a theoretical perspective, it adds to the understanding of how neural networks can implicitly model complex sequential data to provide actionable insights.

Future direction can involve investigating the integration of additional contextual data, optimizing the efficiency of model training and prediction in real-time applications, and broadening the scope of application across diverse processes beyond the datasets studied here. Additionally, further exploration into interpretability and explainability of these models would enhance their adoption in environments where understanding decision rationale is crucial.

By aligning deep learning methodologies with BPM needs, this work significantly contributes to both AI and process management disciplines, suggesting a collaborative path for theory and practice in process-driven prediction models.