Improving event log quality using autoencoders and performing quantitative analysis with conformance checking

Abstract

Event logs work as critical resources for Process Discovery and Conformance along with Enhancement in Process Mining algorithms. Business process event logs, in most practical cases, contain anomalous values as well as missing values which hampers the possibility of extracting useful insights. These logs, when analyzed using the common process mining tools, yield inconsistent results due to the presence of aberrant data values and consequently the business decisions based on their analysis are also of low quality. In order to upgrade the quality of event logs, autoencoders (AE) are used for detecting anomalous values and imputing the missing values. This work is an extension of the methods proposed by Nguyen, Lee, Kim, et al. [11] in their paper tilted ‘Autoencoders for improving quality of process event logs’ where they use different types of autoencoders, and discuss a qualitative analysis of the experiments on the output of process discovery. In this paper, apart from AE and Variational AE (VAE), a GRU based AE is also used and a quantitative analysis of the impact of event log cleaning and reconstruction by performing conformance checking using PM4Py python module for process mining is proposed. The fitness of the original event logs on the process models discovered using the reconstructed event logs, calculated using Token Based Replay as well as Alignments, evaluates to 0.99 and thus clearly justifies the usability of the proposed methods.