Predictive process mining by network of classifiers and clusterers: the PEDF model

TL;DR

This paper introduces the PEDF model, a flexible network of classifiers and clusterers, for predicting future events in process mining, outperforming RNNs and sequential models on event log data.

Contribution

The novel PEDF model combines clustering and classification in a network to improve event prediction accuracy and allows iterative updates, addressing gaps in performance measurement.

Findings

PEDF outperforms RNN and sequential models in event prediction.

The model effectively uses transition differences and cumulative features.

Three new performance measures are proposed for event log prediction.

Abstract

In this research, a model is proposed to learn from event log and predict future events of a system. The proposed PEDF model learns based on events' sequences, durations, and extra features. The PEDF model is built by a network made of standard clusterers and classifiers, and it has high flexibility to update the model iteratively. The model requires to extract two sets of data from log files i.e., transition differences, and cumulative features. The model has one layer of memory which means that each transition is dependent on both the current event and the previous event. To evaluate the performance of the proposed model, it is compared to the Recurrent Neural Network and Sequential Prediction models, and it outperforms them. Since there is missing performance measure for event log prediction models, three measures are proposed.