A. Black-Box Machine Learning for Process Prediction
INTRODUCTION
Business process prediction also referred to as predictive process monitoring or predictive business process management is a branch of process mining that pursues the objective to predict the target of interest by using the activities from the process traces. Recently, several studies have been conducted to explore the applicability of various machine learning approaches for different problems in the process prediction context such as next event prediction, process outcome prediction, prediction of service level agreement violations, remaining time prediction, risk prediction, cost prediction, prediction of activity delays etc. The recent research also suggests that the black-box machine learning approaches especially deep learning methods provide superior results for process prediction problems compared to conventional approaches. However, these opaque, nontransparent models lack the capabilities to provide explanations about their reasoning trace or delivered outcomes. This in turn introduces the barriers to operationalizing data-driven decisionmaking since the users tend not to use the outcomes by such artificial advice givers due to the lack of understanding or justification.
II.
THESIS CONTRIBUTION
reemerged as an important research domain with the purpose to establish the trust between human users and AI systems by making the communication understandable and transparent. Although, the predictive process analytics has been recently emerging as an important research area, the explainability issues in this domain have only been partially addressed. To fill this research gap, this thesis makes three major contributions. First, this study explores the applicability of black-box machine learning approaches, particularly deep neural networks, for different process prediction problems in various domains. Second, this study attempts to propose a theory-driven conceptual framework which is presumed to guide developing explainable process prediction solutions by providing an overview of important aspects of decision-making situations. Third, the applicability of explainable process predictions is illustrated by adopting well-recognized or developing new explanation methods for different use-cases.
The predictive strength of the adopted machine learning
models is one of the most important prerequisites for generating
reliable robust, and consistent explanations. For this purpose, we
have investigated various black box approaches for different
process prediction problems. In one of our earlier studies we
applied pre-trained stacked autoencoder based deep neural
networks to address next event prediction problem after carrying
out an intensive data-preprocessing procedure on the event log
data including n-gram encoding, extracting data flow and
resource features and feature hashing [
Making process predictions delivered by black-box machine
learning models explainable is a multi-dimensional and
multifaceted issue that requires to consider the context of explanation
situation, the users‘ preferences and backgrounds, the nature of
underlying processes, the defined technical and economic
objectives, the organizational factors, etc. when developing
explanation systems. Hence, it conceivable to suggest there is no
“one-fits-all” explainable process prediction solution and a
systematic approach is required in generating adequate
explanations by incorporating the implications from various
dimensions of the decision-making environment. To fill this
research gap, this thesis aims to propose a conceptual
framework which is supposed to guide the process mining
practitioners and researchers in designing and developing the
explanation solutions for predictive process monitoring
problems. For this purpose, our study [
These users pursue various objectives such as justification, ratification, verification, duplication, debugging, learning, satisfaction, effectiveness/efficiency etc. For instance, the knowledge engineers such as process engineers are more interested in the reasoning mechanism of the black-box approaches, follow mainly the verification objective and prefer the explanation methods that provide algorithmic transparency.
On the other side, the domain experts or process owners with limited machine learning background opt to justify the goodness of the models’ individual decisions. There have been numerous studies to investigate various XAI techniques by proposing different taxonomies. E.g. the explanation techniques can be classified as methods related to transparency which aim to examine how the model works or as post-hoc explanation approaches that attempt to extract the explanation from learned model without investigating the reasoning mechanism of the model. Regarding the relationship with the underlying blackbox model the explanations are categorized as model-specific or model-agnostic approaches. A further categorization refers to the scope of the generated explanations. The global explanation approaches attempt to deliver explanations for the whole dataset whereas the local explanation techniques make the individual observations explainable. Generation time informs the users when the explanations are generated. They can be obtained before developing the black-box machine learning models (premodel), during the model implementation (in-model) and finally after the model is trained (post-model).
In this thesis, the proposed conceptual framework is
instantiated to develop various explainable process prediction
solutions. For illustrative purposes, three use-cases are presented
which we have already examined in various studies [
proposed framework for explainable process predictions through the lens of information systems theory, particularly activity theory combined with the design science research. Furthermore, a thorough integration of explanation methods will be examined by ensuring the robust transitions among them with the goal to facilitate the users in examining the explanations from different perspectives. An investigation of xPP for other use-cases such as explaining the reasons for algorithmic fairness violations is another future research direction.