=Paper=
{{Paper
|id=Vol-2973/paper_192
|storemode=property
|title=A Causal Approach to Prescriptive Process Monitoring
|pdfUrl=https://ceur-ws.org/Vol-2973/paper_192.pdf
|volume=Vol-2973
|authors=Zahra Dasht Bozorgi
|dblpUrl=https://dblp.org/rec/conf/bpm/Bozorgi21
}}
==A Causal Approach to Prescriptive Process Monitoring==
https://ceur-ws.org/Vol-2973/paper_192.pdf
A Causal Approach to Prescriptive Process
Monitoring
Zahra Dasht Bozorgi1
1
University of Melbourne, Level 4, 700 Swanston St, Carlton, 3053, VIC, Australia
1. Introduction
A business process is a collection of events, activities, and decisions that collectively lead to an
outcome that can be of value to a customer [1]. Process mining is a family of techniques that
extract information about processes using historical process execution data, generally known
as event logs [2]. Recent advances in process mining techniques have allowed companies to
manage and improve their processes more efficiently. Particularly, process mining has benefited
from advances in machine learning techniques to provide accurate predictions of the future
state of business processes in an area known as predictive process monitoring. Recent predictive
monitoring methods can produce highly accurate predictions of upcoming events, future event
suffixes, remaining time, and the outcome of the process [3, 4, 5, 6, 7, 8]. However, having reliable
predictions does not always lead to improvement of the process. A study by Dees et al. [9], shows
that if good predictions are followed by bad recommendations, the desired improvement is not
achieved. This study illustrates the need for new techniques that find the best interventions
based on the context of each case.
Causal inference is a field in statistics that is concerned with estimating causal effect of a
variable when that variable is changed. This field has been widely used in other domains such
as medicine, social sciences, and marketing. In medicine, causal inference techniques are used
to determine the effectiveness of new drugs. In social sciences and marketing, new policies and
campaigns are evaluated using state-of-the-art causal inference methods. The success of causal
inference in these fields serves as a motivation to use these techniques in the area of process
mining.
In this Ph.D. project, we aim to address the need for effective recommendations by devising
new prescriptive process monitoring methods which are based on causal relationships. These
methods will make recommendations, at tactical and operational levels, about what actions
should be taken to achieve a given process objective.
BPM 2021: Proceedings of the Demonstration & Resources Track, Best BPM Dissertation Award, and Doctoral
Consortium at BPM 2021 co-located with the 19th International Conference on Business Process Management,
September 6-10, 2021, Rome, Italy
" zdashtbozorg@student.unimelb.edu.au (Z. D. Bozorgi)
~ https://www.linkedin.com/in/zahradbozorgi/ (Z. D. Bozorgi)
© 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)
2. Research Problem
Starting from a set of pre-defined treatments (a.k.a. process interventions) and associated costs,
this thesis explores the use of causal inference in prescriptive monitoring, to recommend what
treatments are to be used in what context such that a given benefit function is maximised for
the organisation. Specifically, the overall aim of this project is to answer the following question:
How to provide cost-aware recommendations based on causal relations between a proposed
intervention and a target of interest?
Since the area of prescriptive process monitoring remains largely under-explored, we select
three mains dimensions of exploration for scoping purposes. These are:
• The objective of the recommendations. In this thesis, we focus on two main goals, namely
outcome improvement and remaining time reduction.
• Explainability of the recommendations. We explore the use of both interpretable and
black-box (hard to explain) methods for producing recommendations and assess relative
merits and shortcomings.
• Prescriptive monitoring with multiple treatment options. We explore the use of reinforce-
ment learning solutions to choose the best course of action during the execution of the
process.
Accordingly, we define the following research questions to explore these dimensions:
1. RQ1: How can the use of causal inference in prescriptive process monitoring result
in recommendations that optimise process outcome and cycle time while providing
measurable benefits to the organisation?
2. RQ2: How do black-box approaches for causal inference compare with interpretable
techniques when applied to prescriptive process monitoring?
3. RQ3: How can we use reinforcement learning in combination with causal inference to
devise a prescriptive process monitoring method that can select the best actions that lead
to the achievement of an objective?
3. Approach
To address the identified research problems, we propose three studies each addressing one
research question. To the best of our knowledge, this is the first time causal inference has been
used in a prescriptive process monitoring context.
3.1. RQ1: Recommendations for varying objectives
The main purpose of this study is to investigate the use of causal inference for prescriptive
monitoring. We hypothesise that if we can establish a causal relationship between a possible
intervention and the outcome of interest, the recommendations that result from such causal
relationships will result in more benefit for the company than correlation-based recommenda-
tions. We divided this study into two sub-studies, one addressing outcome improvement and
the other cycle time reduction. We chose this two objectives because one is an example of a
binary outcome, while the other is described as a real-valued attribute.
Our approach proposed in [10] is a rule-based prescriptive system. First, a set of interventions
which are highly correlated with a positive outcome of interest are identified. Then using a
causal effect estimation method called Uplift Tree [11], we identify contexts in which those
interventions causally influence the outcome. We also propose a cost model that identifies
the Return-on-Investment (ROI) of an intervention. The decision to intervene in a process is
then made based on the ROI. We show that using this method, we can automatically identify
improvement actions that are traditionally done by humans.
In another study [12], we propose a cost-aware prescriptive monitoring method that is
designed to reduce cycle time. The core of this recommendation system is the Orthogonal
Random Forest (ORF) [13], which is a causal estimation method that works with continuous
outcome variables. The approach consists of two phases: an offline phase in which given an
event log, a causal effect estimation model is trained, and the best policy for applying the
intervention is selected to maximise the gain in applying it. The second phase is the online
phase where the causal effect estimator and the selected policy are used to determine which
on-going cases should receive an intervention. Our results show that selecting an intervention
policy based on causal models leads to a higher net-gain than policies based on traditional
machine learning methods for prediction.
3.2. RQ2: Explainability of the recommendations
In the second study, we plan to investigate the use of black-box and white-box models. Many
methods for causal estimation have been proposed in the causal inference literature. Most
of these models are black-box, meaning that we cannot understand how different variables
are combined to estimate a causal effect. Conversely, some models such as uplift trees are
interpretable or white-box. These models are constrained to give a better understanding of how
the causal effect estimation is made. Many models, however, are not designed to be interpretable,
but to be accurate estimators of causal effects. Due to an issue known as the fundamental problem
of causal inference, it is very difficult to establish whether black-box models have higher accuracy
in real-world settings and whether the more accurate estimates yield higher benefit. Since
prescriptive monitoring methods are designed to aid humans in decision-making, transparency
of the models they are based on is important. Therefore, in this study, we plan to investigate
whether the use of black-box models is preferable to white-box ones. To do this, we first need to
address the problem of not observing ground truth in the data. So we will first create simulated
event logs by using the data generation approach proposed in [14] based on a flow-based deep
generative model called Sigmoidal Flow. We will train a Sigmoidal Flow on real-life event logs,
and then create simulated data that contain the ground truth causal effect while preserving
the statistical properties of our real-world data. Using this simulated data we will investigate
different black-box and white-box causal inference approaches and develop a cost-benefit model
to determine in which circumstances using a black-box model is warranted.
3.3. RQ3: Multiple treatment options
This study will be conducted in the final year of this Ph.D. In the previous studies, we assume
that there is a binary treatment that can potentially influence a process performance metric.
However, in practice, many actions can be taken during the execution of the process that
influence performance metrics. With the methods proposed in the first study, many causal
models need to be trained separately, and this is computationally inefficient. Furthermore,
using many standalone causal models will not consider how the different actions influence each
other (i.e., the action interplay). In such circumstances, reinforcement learning frameworks
can be employed to design a prescriptive monitoring method that deals with a multitude of
actions/treatments. In particular, contextual bandits algorithms have benefited from the causal
inference literature to make them less prone to problems in estimation bias [15]. Contextual
bandits are an extension of multi-armed bandits. They output an action conditional on the state
of the environment. The aim of this study is to devise a new prescriptive monitoring method
based on contextual bandits that can recommend the best action at each step of the process
execution. In a prescriptive monitoring method based on contextual bandits, the context will
be defined by case prefixes extracted from event logs and ongoing process executions, and the
reward will be a function based on one or multiple process performance metrics.
3.4. Methodology
This project will follow a Design Science research method [16]. The rigour of the approaches
will be ensured by conducting an extensive literature review and constructing a comprehensive
evaluation benchmark, using well defined selection and assessment criteria. The relevance of the
solutions will be ensured via an extensive evaluation of the developed techniques with real-life
and simulated data sets, and where possible, through case studies with relevant organisations.
4. Limitations of the Study
The main limitation of this study is that we assume that all the variables that influence inter-
vention and outcome are observed in the event log. This might lead to biased estimations that
cause the prescriptive system to make a sub-optimal recommendation. We try to alleviate this
threat to validity by performing sensitivity analysis on our models to measure their robustness
to unmeasured confounding. Another limitation of this study is the evaluation of the proposed
methods. While we use cost models and other evaluation methods in the causal inference
literature such as Qini coefficients to evaluate our prescriptive methods, the best and most
rigorous way of evaluating these methods is conducting an A/B test.
References
[1] M. Dumas, M. L. Rosa, J. Mendling, H. A. Reijers, Fundamentals of Business Process
Management, Second Edition, Springer, 2018.
[2] W. M. P. v. d. Aalst, Process Mining: Data Science in Action, second ed., Springer, 2016.
[3] F. Taymouri, M. La Rosa, S. Erfani, Z. D. Bozorgi, I. Verenich, Predictive business pro-
cess monitoring via generative adversarial nets: The case of next event prediction, in:
International Conference on Business Process Management, Springer, 2020.
[4] F. Taymouri, M. L. Rosa, S. M. Erfani, A deep adversarial model for suffix and remaining
time prediction of event sequences, CoRR (2021).
[5] V. Pasquadibisceglie, A. Appice, G. Castellano, D. Malerba, Using convolutional neural
networks for predictive process analytics, in: International Conference on Process Mining,
ICPM 2019, Aachen, Germany, June 24-26, 2019, IEEE, 2019.
[6] N. Tax, I. Verenich, M. L. Rosa, M. Dumas, Predictive business process monitoring with
LSTM neural networks, in: Advanced Information Systems Engineering - 29th International
Conference, CAiSE 2017, Proceedings, Springer, 2017.
[7] M. Camargo, M. Dumas, O. G. Rojas, Learning accurate LSTM models of business processes,
in: T. T. Hildebrandt, B. F. van Dongen, M. Röglinger, J. Mendling (Eds.), Business Process
Management - 17th International Conference, BPM 2019, Vienna, Austria, September 1-6,
2019, Proceedings, Springer, 2019.
[8] L. Lin, L. Wen, J. Wang, Mm-pred: A deep predictive model for multi-attribute event
sequence, in: T. Y. Berger-Wolf, N. V. Chawla (Eds.), Proceedings of the 2019 SIAM
International Conference on Data Mining, SDM 2019, Calgary, Alberta, Canada, May 2-4,
2019, SIAM, 2019.
[9] M. Dees, M. de Leoni, W. M. P. van der Aalst, H. A. Reijers, What if process predictions are
not followed by good recommendations?, in: Proceedings of the Industry Forum at BPM
2019 co-located with 17th International Conference on Business Process Management
(BPM 2019), Vienna, Austria, September 1-6, 2019, CEUR-WS.org, 2019.
[10] Z. D. Bozorgi, I. Teinemaa, M. Dumas, M. L. Rosa, A. Polyvyanyy, Process mining meets
causal machine learning: Discovering causal rules from event logs, in: 2nd International
Conference on Process Mining, ICPM 2020, Padua, Italy, October 4-9, 2020, IEEE, 2020.
[11] P. Rzepakowski, S. Jaroszewicz, Decision trees for uplift modeling with single and multiple
treatments, Knowl. Inf. Syst. (2012).
[12] Z. D. Bozorgi, I. Teinemaa, M. Dumas, M. La Rosa, Prescriptive process monitoring for
cost-aware cycle time reduction, arXiv preprint arXiv:2105.07111 (2021).
[13] M. Oprescu, V. Syrgkanis, Z. S. Wu, Orthogonal random forest for causal inference, in:
Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15
June 2019, Long Beach, California, USA, PMLR, 2019.
[14] B. Neal, C.-W. Huang, S. Raghupathi, Realcause: Realistic causal inference benchmarking,
arXiv preprint arXiv:2011.15007 (2020).
[15] M. Dimakopoulou, Z. Zhou, S. Athey, G. Imbens, Balanced linear contextual bandits, in:
Proceedings of the AAAI Conference on Artificial Intelligence, 2019.
[16] R. H. von Alan, S. T. March, J. Park, S. Ram, Design science in information systems research,
MIS quarterly (2004).