Event Log Reconstruction Using Autoencoders

Abstract

Poor quality of process event logs prevents high quality business process analysis and improvement. Process event logs quality decreases because of missing attribute values or after incorrect or irrelevant attribute values are identified and removed. Reconstructing a correct value for these missing attributes is likely to increase the quality of event log-based process analyses. Traditional statistical reconstruction methods work poorly with event logs, because of the complex interrelations among attributes, events and cases. Machine learning approaches appear more suitable in this context, since they can learn complex models of event logs through training. This paper proposes a method for reconstructing missing attribute values in event logs based on the use of autoencoders. Autoencoders are a class of feed-forward neural networks that reconstruct their own input after having learnt a model of its latent distribution. They suit problems of unsupervised learning, such as the one considered in this paper. When reconstructing missing attribute values in an event log, in fact, one cannot assume that a training set with true labels is available for model training. The proposed method is evaluated on two real event logs against baseline methods commonly used in the literature for imputing missing values in large datasets.