Process mining with streaming data
Year: 2019
Type: dissertation
Abstract: Modern information systems allow us to track, often in great detail, the execution of processes within companies.Consider for example luggage handling in airports, manufacturing processes of products and goods, or processes related to service provision, all of these processes generate traces of valuable event data.Such event data are typically stored in a company's information system and describe the execution of the process at hand.In recent years, the field of process mining has emerged.Process mining techniques aim to translate the data captured during the process execution, i.e. the event data, into actionable insights.As such, we identify three main process mining types of analysis, i.e. process discovery, conformance checking and process enhancement.In process discovery, we aim to discover a process model, i.e. a formal behavioural description, which describes the process as captured by the event data.In conformance checking, we aim to assess to what degree the event data is in correspondence with a given reference model, i.e. a model describing how the process ought to be executed.Finally, within process enhancement, the main goal is to improve the view of the process, i.e. by enhancing process models on the basis of facts derived from event data.Recent developments in information technology allow us to capture data at increasing rates, yielding enormous volumes of data, both in terms of size and velocity.In the context of process mining, this relates to the advent of real-time, online, streams of events that result in data sets that are no longer efficiently analysable by commodity hardware.Such types of data pose both opportunities and challenges.On the one hand, it allows us to get actionable insights into the process, at the moment it is being executed.On the other hand, conventional process mining techniques do not allow us to gain these insights, as they are not designed to cope with such a new type of data.As a consequence, new methods, techniques and tools are needed to allow us to apply process mining techniques and analyses on streams of event data of arbitrary size.In this thesis, we explore, develop and analyse process mining techniques that are able to handle streaming event data.The premise of streaming event data, is the fact that we assume the stream of events under consideration to be of infinite size.As such, efficient techniques to temporarily store and use relevant recent subsets of event data vi are needed.The techniques developed in the context of this thesis allow us to apply process mining techniques using potentially unbounded streams of data with arbitrary rates of emission.The ability to handle such data allows us to analyse the underlying process at the exact moment it is being executed.Such analysis paves the way for more advanced types of process mining techniques such as real-time process monitoring and real-time prediction.Since the techniques developed are able to handle data of arbitrary size, as a side-effect, they allow us to handle data sets that are beyond the size-limitation of conventional process mining techniques.The contributions of this thesis can be categorized into four separate dimensions, all having a strong link with the main branches of process mining.1.The development of techniques for data storage and data quality.We provide a general formalization for temporal storage of event data.We furthermore show that we are able to instantiate the proposed formalization using a vast array of existing data storage techniques.As such, we are able to lift all conventional process mining technique to the domain of streaming data.We furthermore present means to filter streaming event data, which allows us to increase the overall data quality considered.2. The development of techniques for event stream based process discovery.We explicitly lift process discovery to the domain of event streams by means of designing a general purpose architecture, which describes a two-step discovery approach.The first step consists of constructing an algorithm-specific intermediate data representation, on the basis of the event stream, whereas the second step consists of translating the intermediate representation to a process model.We furthermore show that the proposed architecture covers a wide variety of process discovery algorithms.We additionally show advanced results for the class of language-based region theory-based process discovery algorithms, where we primarily focus on exploiting the intermediate representation of the algorithm to further improve process discovery results. The development of techniques for event stream based conformance checking.We propose a greedy computational approach for the purpose of computing conformance checking statistics on the basis of event streams.We furthermore prove that the approach is to be seen as an under-estimator, which implies that when a deviation is observed, we are guaranteed that something went wrong during the execution of the process. The development of techniques for event stream based process enhancement.Next to the control-flow perspective, we also provide support for the resource perspective.In particular, we examine the suitability of a set of existing resource network metrics in the context of event streams.For each of the aforementioned contributions, a corresponding prototypical implementation is provided in both the process mining tool-kits ProM (http://promtools.vii org) and RapidProM (http://rapidprom.org).As such, each technique discussed has an accompanying publicly available implementation that is used within the corresponding evaluation. IntroductionModern information systems allow us to track, often in great detail, the execution of processes within companies.Examples of such processes concern luggage handling in airports, manufacturing processes of products and goods and processes related to service provision.All of these processes generate traces of valuable event data.Such event data are typically stored in a company's information system and describe the execution of instances of the process at hand.Process mining [4] is a relatively young, purely data-driven, research discipline within computer science that is positioned in-between traditional data mining [15] on the one hand and business process management [51] on the other hand.The main goal of process mining is to gain insights in, and knowledge of, the behaviour of the processes executed within a company.In particular, we aim to attain such knowledge based on the event data that is generated during the execution of the process and stored in a company's underlying information system.As such, the diagrams and process models obtained by the application of process mining, represent, under the assumption that event data is recorded correctly, what actually happened during execution of the process.Within process mining, we distinguish three main types of analysis, i.e. process discovery, conformance checking and process enhancement.These types of analysis mainly differ in the types of input elements they require, as well as their intended analysis result.In process discovery, the main goal is to discover a process model describing the process under study, based on the behaviour captured within the event data.In conformance checking, the main goal is to verify to what degree a given process model (possibly discovered) and the process under study, again as captured by the corresponding event data, conform to one another.In process enhancement, the main goal is to improve the overall view of the process by improving/extending a process model based on facts and figures deduced from behaviour captured within the event data, e.g. by adding performance statistics such as bottleneck information.Due to the ever increasing performance of computational systems and architectures, data are being generated at increasingly high rates, yielding data that are challenging in terms of volume, velocity, variety and veracity.For example, consider the fact that, with the rise of the use of mobile phones and their connection to the internet, virtually Process enhancement in the context of this thesis In this thesis, in chapter 8, we present a study towards computing social networks, i.e. networks of resources cooperating together during process execution, in the context of event streams.We show that there exist variations of such networks that lend themselves for incremental computation, which in turn allows us to adopt these variations in an online, event stream setting. Learning from Data StreamsIn this thesis, we primarily focus on the application of process mining techniques in the context of streaming data, i.e. infinite streams of data elements that potentially arrive at unprecedented velocity.Such streaming data, as well as its corresponding formal model, are well-studied concepts in the field of data mining.Based on [64, 95], we define the general data stream model as follows.We receive a sequence of input data elements d 1 , d 2 , ..., d i , ..., that describe some, typically unknown, underlying generating function D. In general, two types of streaming models are recognized, i.e.• Insertion-Only/Cash Register Once a data item d i is received, it is not changed, i.e. it remains as-is.• Implicit Start/End Apart from data incompleteness, we assume that we have no explicit knowledge of process instance initialization and termination.As such, it is not clear when a process instance started and when it is safe to assume its termination.This implies that, when we observe an event on the event stream related to a case identifier that was not observed before, we have no guarantee that such an event is indeed the first event of the underlying process instance.Moreover, We develop techniques that go beyond efficient event storage and decrease the memory consumption need for the purpose of process discovery.• Research Goal III Development/design of specialized techniques for efficient,
Source: TU/e Research Portal
Author: Sebastiaan J. van Zelst
Institution: Eindhoven University of Technology
Language: English
Subfield: Management Information Systems
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W2920149474
