Applicability of Process Mining Techniques in Business Environments

This thesis analyses problems related to the applicability, in business environments, of Process Mining tools and techniques.

The first contribution is a presentation of the state of the art of Process Mining and a characterization of companies, in terms of their "process awareness". The work continues identifying circumstance where problems can emerge: data preparation; actual mining; and results interpretation. Other problems are the configuration of parameters by not-expert users and computational complexity.

We concentrate on two possible scenarios: "batch" and "on-line" Process Mining.

Concerning the batch Process Mining, we first investigated the data preparation problem and we proposed a solution for the identification of the "case-ids" whenever this field is not explicitly indicated.
After that, we concentrated on problems at mining time and we propose the generalization of a well-known control-flow discovery algorithm in order to exploit non instantaneous events. The usage of interval-based recording leads to an important improvement of performance.
Later on, we report our work on the parameters configuration for not-expert users. We present two approaches to select the "best" parameters configuration: one is completely autonomous; the other requires human interaction to navigate a hierarchy of candidate models.
Concerning the data interpretation and results evaluation, we propose two metrics: a model-to-model and a model-to-log. Finally, we present an automatic approach for the extension of a control-flow model with social information, in order to simplify the analysis of these perspectives.

The second part of this thesis deals with control-flow discovery algorithms in on-line settings. We propose a formal definition of the problem, and two baseline approaches. The actual mining algorithms proposed are two: the first is the adaptation, to the control-flow discovery problem, of a frequency counting algorithm; the second constitutes a framework of models which can be used for different kinds of streams (stationary versus evolving).

Thesis available at http://andrea.burattin.net/publications/2013-phd-thesis

Transcript

1. Applicability of Process Mining Techniques in Business Environments Candidate Andrea Burattin, XXV Ciclo Supervisor Prof. Alessandro Sperduti April 8, 2013
2. 2 of 45 Table of Contents  Information systems and business processes  Process mining  Application of process mining in business and industrial environments – Problems – Possible solutions  Conclusions and future work
3. 3 of 45 Current Information Systems  Information systems – Usage of information systems growing in all companies – Information systems moving from single “vertical” functionalities towards “horizontal” business processes – All these systems record many data but its exploitation, typically, is not straightforward
4. 4 of 45 Business Processes  There is no comprehensively accepted definition of “business process”, however, shared features are – A finite set of activities (and dependencies) – Activity performers or originators – Output produced as execution result  Typical example
5. 5 of 45 Event Logs  Executions traces of business processes are typically recorded in log files
6. 6 of 45 What is Process Mining?  Input – An event log – Optionally, other a-priori knowledge (e.g., a process model)  Examples of possible outputs – A model describing how activities are performed • Useful to compare ideal process versus the actual one – Relationships among originator • Useful to redistribute resources over activities – Statistics on the performance of the execution • Useful for monitoring purposes
7. 7 of 45 Process Mining Overview Image source: Christian W. Gűnther. Process mining in Flexible Environments. PhD thesis, Technische Universiteit Eindhoven, Eindhoven, 2009 .
8. 8 of 45 Theoretical Open Problems  Process mining literature presents several open problems, for example (van der Aalst, Comp. in Ind., 2004) – Duplicate tasks • Activities with the same name in different positions of the model – Exploiting all data available • For example, not all the algorithms use all the time information to distinguish the starting from the finishing time of an event – Holistic mining • Different perspectives from different sources: not only the control flow but also other perspectives, in order to create a global process description – Noise and incompleteness • Obtaining a complete log, where all the required information are actually available
9. 9 of 45 Industrial-related Open Problems  Our case studies revealed other problems – Using process mining tools and configuring algorithms – Results interpretation • Generation of the results with an as-readable-as-possible [graphical] representation of the process: information are represented in a simple and understandable manner – Computational power and storage capacity required • Small and medium sized companies may not be able to cope with the technological requirement of large process mining projects
10. 10 of 45 Possible Industry Scenarios  We characterized four possible scenarios, based on process awareness of companies and of their systems – Company process aware vs. Process unaware – Process aware software vs. Process unaware software
11. 11 of 45 Thesis Structure and Organization
12. 12 of 45 Overview – Data Preparation
13. 13 of 45 Problems with Data Preparation  Several problems with data preparation, at different complexity and abstraction levels  Key points – Adaptation of existing data (“syntax problem”, easy) – Construction of all the required information – Introduction of new information
14. 14 of 45 Problems with Data Preparation – 2  Process mining required fields are (activity; process-name; case-id; timestamp; originator)  In our real case, we have – Company process aware – Information systems process unaware – Log structured as (activity; timestamp; originator; info1; …; infon)
15. 15 of 45 Problems with Data Preparation – 3  The name of the process is not a problem (we can assume all events belonging to the same process)  To extract the case-id from info fields (idea) – Isolation of candidate case-id fields (a-priori knowledge) – Construction of “event chains”, binding two events that share at least one field's value (strings similarity functions) – Selection of the maximal chain (one with most activities or the simplest chain)  Details reported in (Burattin & Vigo, CIDM, 2011)
16. 16 of 45 Overview – Control-flow Mining
17. 17 of 45 Basic Idea of Control-flow Discovery  Basic idea of control-flow discovery algorithms is the identification of dependencies between activities  If the event log contains sequences as – … A, B … Notation: A > B – … B, A …  Discovered model contains dependency from A to B: A B
18. 18 of 45 Exploiting Data Available  Events with duration instead of instantaneous event  Generalization (i.e. same parameters) of a currently available control-flow mining algorithm to exploit this new information – Dependencies for time intervals  Dependency measure  AND-measure
19. 19 of 45 Exploiting Data Available – 2  Evaluation of new algorithm against synthetic dataset (100 processes, 1000 cases, 12000 events, 10% noise)  Test against “real” dataset, almost correct mining  Reported in (Burattin & Sperduti, ESANN, 2010), implemented in ProM 5.2
20. 20 of 45 Not-expert Users  Typical users of process mining are “not-expert” users  Not-expert users in process mining but we assume they have notions in process modelling – Process mining algorithms provide configurations (parameters) to cope with different scenarios • Heuristics Miner parameters are thresholds on measures (e.g., dependency measure and AND-measure) – Process mining algorithms are implemented in tools – Not-expert users don't understand algorithm and are not skilled in using tools
21. 21 of 45 Basic Idea  We defined an approach to discretize the values of each parameters of the algorithm we are considering – Idea: the log is finite, therefore only a finite number of significant thresholds exist  Each parameter configuration defines a model  We can shift the problem of configuring parameters of process mining algorithms into “selecting the best model out of a set” – Automatic approach to explore this space – User-guided approach
22. 22 of 45 Automatic Discovery  Hill climbing with maximum plateau steps to reach local optimum  Several random restarts to try to achieve the global optimum  Optimization criterion based on Minimum Description Length principle – MDL encoding for process mining (Calders et al., SAC, 2009) – Heuristics to improve efficiency and performance
23. 23 of 45 Automatic Discovery – Results  Tests against synthetic dataset: 93 process models, 250 cases for each process – Rigorous adaptation with 3 values for α (balancing the model complexity and the data explanation)
24. 24 of 45 Automatic Discovery – Results 2  Performance in terms of time required to process different models, with the two techniques  Reported in (Burattin & Sperduti, CEC, 2010), implemented in ProM 5.2
25. 25 of 45 User-guided Discovery  Model comparison using a model-to-model metric (Burattin et al., BPI, 2011)  Approach based on hierarchical clustering with average linkage  In our implementation, user can “explore” the hierarchy
26. 26 of 45 Overview – Results Evaluation
27. 27 of 45 Process Mining Evaluation  It is possible to evaluate results of process mining using – Model-to-model metrics: to compare two given process models (typically the target and the mined one) • Structural Similarity • Dependency Difference Metric – Model-to-log metrics: to define the amount of behaviour observed in the log that is allowed by the model • Fitness • Soundness
28. 28 of 45 Model-to-model Metric  Metric used to cluster processes (Burattin et al., BPI, 2011) – Decomposition of complex process models into • Allowed / Forbidden behaviour – Comparison using Jaccard Similarity
29. 29 of 45 Model-to-log Metric  Process model as a set of Declare constraints  Given a Declare constraint π and a trace σ  Healthiness measures as – Activation sparsity – Violation ratio – Fulfilment ratio – Conflict ratio  Reported in (Burattin et al., EDOC, 2012), implemented in ProM 6.2
30. 30 of 45 Overview – Process Extension
31. 31 of 45 Multi-perspective Mining  Extension of control-flow model with social perspective  Input – Log with information on originators – Process model  Output – Process model extended with roles
32. 32 of 45 Multi-perspective Mining – 2  Each role requires a different set of skills – We assume that skills are characterized by set of users  Basic idea is to identify handover of roles starting from actual process dependencies  We don't have handover of role between a and b if – In some cases a and b are performed by the same originator – If the set of originators performing a and b are similar  We divide activities in roles according to threshold τw
33. 33 of 45 Multi-perspective Mining – 3  Identifying handover is not sufficient  Sometimes we need to merge roles  We merge roles according to a threshold τρ
34. 34 of 45 Multi-perspective Mining – 4  We are able to discretize τw and τρ to their significant values  We use an entropy-based metric to rank the most interesting partitions  Reported in (Burattin et al., CIDM, 2013), impl. in ProM 6.2
35. 35 of 45 Overview – Stream Control-flow Mining
36. 36 of 45 Basic Idea
37. 37 of 45 Stream Mining Peculiarities  Peculiarities of the stream mining problem – It is impossible to store the complete stream – Backtracking over a data stream is not feasible, so algorithms are required to make only one pass over data – It is important to quickly adapt the model to cope with unusual data values – The approach must deal with variable system conditions, such as fluctuating stream rates  Basic idea is that recent observations are more important than older ones
38. 38 of 45 Proposed Approaches  Three approaches for stream process mining – A basic approach with two variations (sliding window and periodic reset) • Parameter: maximum space occupied – The adaptation of an approach for approximate frequency counting (i.e. Lossy Counting) to our problem • Parameter: maximum error allowed – A new general framework specifically design for process mining which support several update policies • Parameters: maximum space occupied, weighting policy
39. 39 of 45 Weights Update Policies  The framework is based on the continuous update of frequencies of Heuristics Miner basic relations – Most frequent activities – Most frequent direct succession relations  Policy for Stationary Streams – Online Heuristics Miner (with error bounds)  Policies for Evolving Streams – Online Heuristics Miner with Ageing – Online Heuristics Miner with Self-Adapting Ageing
40. 40 of 45 Stream Miner Results  Tests against 3 realistic datasets, and one real dataset  Results on a dataset with 6000 cases and 58783 events  Results in technical report (CoRR abs/1212.6383), implemented in ProM 6.2
41. 41 of 45 Stream Miner Results – 2
42. 42 of 45 Overview
43. 43 of 45 Processes and Logs Generator  Many companies are reluctant to share their data  Stochastic context free grammar to generate random business processes – Rules to simulate a process in order to produce an event log – Reference model and log can be used for the evaluation of control-flow mining algorithms  Reported in (Burattin & Sperduti, BPI, 2011)  Implemented in independent software, now part of official ProM 6.2 distribution
44. 44 of 45 Conclusions  Process mining techniques may have an important impact on companies work (especially with SMEs)  Several problems prevent the application of process mining techniques in real business environment – Data availability – Actual mining phase (both in mining algorithms and in human interaction) – Results interpretation  Solutions have been proposed on all the stages where problems may arise
45. 45 of 45 Detailed Map of Performed Activities