Learning Accurate LSTM Models of Business Processes

Camargo, Manuel; Dumas, Marlon; González-Rojas, Oscar

doi:10.1007/978-3-030-26619-6_19

Cited by 149 publications

(250 citation statements)

References 12 publications

Supporting

Mentioning

243

Contrasting

Unclassified

Order By: Relevance

“…Sequence prediction is concerned with predicting future events based on past events and can be applied to both ongoing and completed cases of an event log. Next event predictions produced by DL models have been demonstrated to be more accurate than those made through process models discovered using event logs [11], [21], [10]. However, the decision-making process of how these predictions are made remains unclear when using DL models.…”

Section: Process Discovery Versus Process Prediction Methodsmentioning

confidence: 99%

“…However, the models did not perform well when making long-term predictions and when used on event logs with many repeated events. To overcome these limitations, Camargo et al [21] proposed an LSTM architecture with embedded dimensions to predict traces of events, time-stamps and the role associated with each event. The phases in their approach included scaling and extracting n-grams of fixed sizes for each event log trace to create input and target sequences for training the predictive model, and a post-processing phase, where predicted next events were randomly selected from the likely ones to generate a greater number of different traces.…”

Section: ) Deep Learning In Process Miningmentioning

confidence: 99%

“…The output layer uses the softmax activation function to ensure the output takes the form of probability distributions. While some of the existing LSTM models for process mining add a dense layer on top of the LSTM layer, they either do not use embedding at all or implement it differently (e.g., see [20] and [21]). Furthermore, the existing LSTM models for process mining employs two LSTM layers, whereas the proposed models have only one LSTM layer.…”

Section: ) Encoding and Paddingmentioning

confidence: 99%

“…The details of these logs are provided in the next section and Table 4. The accuracy of the next event predictions of the proposed two models was compared to that of three other LSTM models proposed by Tax et al [20], Camargo et al [21] and Lin et al [23]. From the five real-life logs listed above, these studies employed only the Helpdesk and BPIC 2012 logs.…”

Section: A Evaluationmentioning

confidence: 99%

“…Tax et al [20] used one-hot encoding to encode events as inputs to an LSTM model that had two layers and 100 neurons in each layer. Camargo et al [21] extracted n-grams of fixed sizes for each event log trace, which they used as inputs to their LSTM model consisting of two stacked LSTM layers and a dense output layer. Also, Camargo et al [21] used pretrained embedded dimensions in all their experiments.…”

Section: A Evaluationmentioning

confidence: 99%

See 4 more Smart Citations

A Graph-Based Approach to Interpreting Recurrent Neural Networks in Process Mining

2020

View full text Add to dashboard Cite

Process mining is often used by organisations to audit their business processes and improve their services and customer relations. Indeed, process execution (or event) logs constantly generated through various information systems can be employed to derive valuable insights about business operations. Compared to traditional process mining techniques such as Petri nets and the Business Process Model and Notation (BPMN), deep learning methods such as Recurrent Neural Networks, and Long Short-Term Memory (LSTM) in particular, have proven to achieve a better performance in terms of accuracy and generalising ability when predicting next events in business processes. However, unlike the traditional network-based process mining techniques that can be used to visually present the entire discovered process, the existing deep learning-based methods for process mining lack a mechanism explaining how the predictions of next events are made. This study proposes a new approach to process mining by combining the benefits of the earlier, visually explainable graph-based methods and later, more accurate but unexplainable deep learning methods. According to the proposed approach, an LSTM model is employed first to find probabilities for each known event to appear in the process next. These probabilities are then used to generate a visually interpretable process model graph that represents the decision-making process of the LSTM model. The level of detail in this graph can be adjusted using a probability threshold, allowing to address a range of process mining tasks such as business process discovery and conformance checking. The advantages of the proposed approach over existing LSTM-based process mining methods in terms of both accuracy and explainability are demonstrated using real-world event logs.

show abstract

Section: Process Discovery Versus Process Prediction Methodsmentioning

confidence: 99%

Section: ) Deep Learning In Process Miningmentioning

confidence: 99%