Explainable concept drift in process mining

Highlights

• We propose a data-driven approach to uncover explanations for process concept drifts.

• The proposed technique detects and explains drifts for object-centric event data.

• Our framework supports linear and non-linear relationships.

Abstract

The execution of processes leaves trails of event data in information systems. These event data are analyzed to generate insights and improvements for the underlying process. However, companies do not execute these processes in a vacuum. The fast pace of technological development, constantly changing market environments, and fast consumer responses expose companies to high levels of uncertainty. This uncertainty often manifests itself in significant changes in the executed processes. Such significant changes are called concept drifts. Transparency about concept drifts is crucial to respond quickly and adequately, limiting the potentially negative impact of such drifts. Three types of knowledge are of interest to a process owner: When did a drift occur, what happened, and why did it happen. This paper introduces a framework to extract concept drifts and their potential root causes from event data. We extract time series describing process measures, detect concept drifts, and test these drifts for correlation. This framework generalizes existing work such that object-centric event data with multiple case notions, non-linear relationships, and an arbitrary number of process measures are supported. We provide an extendable implementation and evaluate our framework concerning the sensitivity of the time series construction and scalability of cause–effect testing. Furthermore, we provide a case study uncovering an explainable concept drift.

Introduction

Throughout the past decades, information systems have become an elementary component supporting various businesses. Executions of business processes [1] leave traces of event data in such information systems, describing the conducted actions. These data can be analyzed to generate insights and improvements for the supported processes. The techniques dealing with these types of problems are summarized under the term process mining [2]. Applying such techniques helps businesses increase efficiency by understanding their processes and is, therefore, a fundamental part of every competitive company.

Different process mining techniques have different objectives of knowledge discovery. Process discovery, for example, delivers a visual representation of the possible executions paths within a process [3]. Conformance checking techniques allow a user to uncover deviations of process executions and measure the correspondence of a process to the recorded event data [4]. Other techniques aim to identify and solve problems within a process, e.g., by simulation [5] or by predicting problematic process executions [6].

Concept drift is one process-related problem that has drawn attention over the past years [7], [8], [9]. A concept drift is a significant change in a process over time. The occurrence of a concept drift may lead to several problems or inefficiencies. However, concept drifts are challenging to detect and analyze because the process is already dynamic, exhibiting stochastic behavior. In the literature, three main questions about problems related to concept drifts are of importance: (1) When did the concept drift occur? If a concept drift happens in the control-flow of the process, i.e., the ordering of executed actions, process discovery, conformance checking, and predictive monitoring techniques have to be adjusted accordingly [10]. (2) What happened and how did it happen? Significant changes need to be understood by the process owner to react accordingly [11]. (3) Why did it happen? Significant changes might be caused by other changes in the process [12]. Uncovering potential cause–effects helps to uncover problems and improve the process.

These three questions are the main research lines related to concept drift in process mining. Different works aim to solve one or multiple of these questions. In this work, we focus on the third research question. However, we can also generate insights for the first two research questions as a byproduct of our work.

• Detection The existence of a concept drift needs to be detected, and its change point needs to be located as precisely as possible.

• Characterization The nature of the concept drift needs to be described as accurately and extensively as possible.

• Explanation Potential cause–effects of a concept drift contained within the event data need to be uncovered.

Several techniques have addressed RQ1 over the last years, with a focus on solving this problem for the control-flow of a process [13]. Generally, these techniques first calculate some numeric representation of the control-flow and, subsequently, use one of many change point detection algorithms [14] such as hypothesis testing, cost-based segmentation techniques, or visual inspection. Different techniques have been proposed to answer RQ2, both in an interactive and fully automated manner. Yeshchenko et al. [9] provide an extensive visual analytics framework for humans to explore concept drifts and understand the dynamics and changes behind them. Ostovar et al. [15] uncover change patterns of a concept drift by testing the data against predefined business process change patterns. These are given as textual output to the user. In recent work, we propose a general framework to answer RQ3 [12]. In this paper, we generalize and extend our previous work to address several challenges encountered in real-life information systems.

The first of these challenges is the so-called “object-centricity” of information systems: Traditional process mining techniques assume the existence of a single case notion and that each recorded event is associated with exactly one object of the case notion. In reality, an information system, e.g., an ERP system, consists of many case notions, e.g., different document types, and events may be related to multiple objects of different case notions [16]. To apply traditional process mining techniques, these object-centric event data have to be flattened first, forcing them into traditional event log format [17]. Flattening is related to certain problems (cf. Section 4.1) and provides misleading insights. Therefore, process mining techniques have to be adapted into the object-centric setting to provide accurate insights. Recently, academia and industry have picked up this challenge. On the one hand, many techniques have been proposed to translate traditional process mining problems to the object-centric setting [18], [19], [20], [21], [22], [23], [24]. On the other hand, industry leaders provide initial support of object-centric event data, e.g., Celonis by supporting object-centric process models through ProcessSphere, or Mehrwerk Process Mining by supporting multiple case identifiers.

The second challenge is non-linear cause–effect relationships contained in information systems. The investigation of cause–effect relationships behind concept drifts cannot be limited to linear relationships. Different dynamics in processes, e.g., workload–productivity relationships of resources [25], may show non-linear behavior.

The third challenge is the absence of domain knowledge. One cannot always assume that a basic knowledge or suspicion of a candidate perspective for a concept drift and its potential cause is present. In our original framework, the user has to choose one perspective to be investigated for concept drifts and another perspective to be investigated for potential causes. However, an approach stripped from its necessity for domain knowledge would be more generally applicable.

Therefore, the work presented in this paper is a generalization and extension of our original framework [12] in the following ways: (a) Our revised framework supports event data with multiple case notions. (b) We support the detection of non-linear relationships. (c) No choice about a primary and secondary perspective has to be made. The user can choose arbitrarily many features.

Our new framework is depicted in Fig. 1. The event log is first segmented into subsequent windows. For each of these windows, we calculate multiple numerical features subject to the user’s choice. These values are concatenated into a time series for each feature. Subsequently, we detect concept drifts in these time series. For each pair of features and each pair of concept drifts of the two series, we test for Granger causality [26] given the time difference between drifts. Non-linear relationships are covered by applying a kernel function. Granger-causal concept drift pairs are given to the user as explainable concept drifts.

We answer RQ1 by detecting concept drifts using existing concept drift detection techniques. The time series provide a visualization to explore the nature of the concept drift, helping in answering RQ2. The correlated concept drifts give explanations and potential root causes for drifts, answering RQ3.

The remainder of this paper is structured as follows: Section 2 introduces related work on concept drift in process mining. We formalize object-centric event data in Section 3. Extracting time series from object-centric event logs is introduced in Section 4. We give a general definition for concept drift detection in Section 5. Section 6 introduces Granger causality for time series and testing for non-linear relationships. Our general framework and a short overview of the implementation is given in Section 7. In Section 8, the framework is, first, evaluated for sensitivity and scalability and, second, applied to a real-life event log uncovering an explainable concept drift in a case study. We conclude this paper in Section 9.