How LIME Explanation Models Can Be Used to Extend Business Process

Models by Relevant Process Details

Myriel Fichtner

, Stefan Sch

onig

and Stefan Jablonski

University of Bayreuth, Germany

University of Regensburg, Germany

Keywords:

Image Mining, Business Process Model Enhancement, Business Process Improvement, Relevant Process

Detail, Convolutional Neural Network, LIME Explanation Models.

Abstract:

Business process modeling is an established method to describe workﬂows in enterprises. The resulting models

contain tasks that are executed by process participants. If the descriptions of such tasks are too abstract or do

not contain all relevant details of a business process, deviating process executions may be observed. This leads

to reduced process success regarding different criteria, e.g., product quality. Existing improvement approaches

are not able to identify missing details in process models that have an impact on the overall process success.

In this work, we present an approach to extract relevant process details from image data. Deep learning

techniques are used to predict the success of process executions. We use LIME explanation models to extract

relevant features and values that are related to positive process predictions. We show how a general conclusion

of these explanations can be derived by applying further image mining techniques. We extensively evaluate

our approach by experiments and demonstrate the extension of an existing process model by identiﬁed details.

1 INTRODUCTION

Business process modeling languages are used to de-

scribe business processes. Business process models

provide insights into the structure of processes and

potential for process improvement. Usually, process

models are designed by process experts and thus con-

tain all information that have been identiﬁed as impor-

tant by experience. However, the results of process

model executions might not be optimal. It may even

be observed that the execution of the same process

model results in different outcomes that differ with

respect to various criteria, e.g., production time, qual-

ity or cost of the process output. Such observations

lead to the assumption that the process is not modeled

in sufﬁcient detail or quality. In this research we ex-

plicitly focus on missing details of process models as

cause for non-optimal executions. We identify three

reasons why a process model is lacking details.

(i) Process modelers are often not sure how de-

tailed a process has to be modeled in order to guar-

antee optimal executions. Although, there are a lot

of modeling tools available, none of them provides

modeling guidelines in this sense (Kluza et al., 2013).

There are styling guidelines and modeling advice in

order to keep aspects like clarity (e.g., (Becker et al.,

2000), (Mendling et al., 2010)), but there is no in-

struction how detailed a process should be designed.

Even process experts are not aware of all details or

which details are relevant enough to model (Nieder-

mann et al., 2010). Input constellations or prerequi-

sites are taken as granted and are not considered in

process descriptions. Due to lack of awareness, im-

portant details are missing in a process model.

(ii) Modelers know which details are relevant for

process success, but they cannot integrate them ap-

propriately into a model. Consider a task that allows

multiple input constellations, but not all of them lead

to successful process executions. This information is

hard to represent in a process model. In such cases,

the restricted expressiveness of established modeling

languages leads to the lack of modeling details. To ex-

tend the usability and expressiveness of existing mod-

eling languages, different approaches have been de-

veloped. For example, media data can be attached to

modeling elements (Wiedmann, 2017). Nevertheless,

the relevant information is missing in the model as

long as no suitable representation method is available.

(iii) Some information can be intentionally omit-

ted in the process model. Complex processes lead

to complex models which contain hundreds of mod-

eling elements. Large process models, in turn, lead

Fichtner, M., Schönig, S. and Jablonski, S.

How LIME Explanation Models Can Be Used to Extend Business Process Models by Relevant Process Details.

DOI: 10.5220/0011067600003179

In Proceedings of the 24th International Conference on Enterprise Information Systems (ICEIS 2022) - Volume 2, pages 527-534

ISBN: 978-989-758-569-2; ISSN: 2184-4992

527

to overload. In the worst case, process participants

are unable to read and execute the tasks appropriately.

Therefore, one goal of modeling is to avoid too com-

plex models. This requirement is hard to meet, while

process designers try to reduce the size of models by

(a) omitting details during modeling that are known

but do not occur often enough or (b) merging alterna-

tive cases due to their presumed irrelevance or rarity.

By using abstraction mechanisms (e.g., (Polyvyanyy

et al., 2008), (Reichert et al., 2012)), modeling ele-

ments are aggregated leading to coarser views that

support comprehensibility. Then, process informa-

tion is omitted for the beneﬁt of abstraction. The loss

of information may include relevant details that inﬂu-

ence process success.

These reasons show that non-modeled but relevant

process details have to be identiﬁed. It is not enough

to determine missing aspects, but also to identify their

characteristics that optimize process output. Existing

process improvement approaches do not cover these

requirements, since they are based on pre-known as-

pects. Most of them focus on how to optimally re-

structure already deﬁned modeling elements (e.g., ad-

justment of execution paths). However, it is indeed

possible that important process details are not taken

into account during the modeling stage and thus can-

not be analyzed by such techniques. Examples can be

found in manufacturing environments, where many

tasks are executed manually by process participants.

In industrial projects we face placement scenarios

where process models just prescribe that parts should

be placed on a pallet. Further instructions on how to

place the parts are missing. Observing process execu-

tions reveal that the process outcomes are deviating.

Our research goal is motivated by this experience.

Existing process improvement approaches show

that additional data sources and analysis concepts

have to be taken into account. In manufacturing envi-

ronments, production lines are usually monitored by

collecting image data. According to (Schmidt et al.,

2016), images are a powerful data source, which con-

tain complex information and interrelations, where

business processes can beneﬁt from. Inspired by this,

we aim for the deployment of image mining tech-

niques. In this paper we focus on the implementation

of an approach that extracts relevant process details

from images to enhance process models and guar-

antee process success. We propose to collect image

data of single task executions and to analyze them us-

ing LIME explanation models (Ribeiro et al., 2016).

Based on these results, insufﬁcient modeled details

are identiﬁed. Examples on how these details may be

attached to a process model are given. Our work com-

plements previous process improvement approaches.

It supports process designers even after the modeling

phase in their efforts of designing a process success-

oriented model.

2 RELATED WORK

Business processes describe a series of steps that must

be performed to achieve a business goal, such as

manufacturing a speciﬁc product. The modeling lan-

guage Business Process Model and Notation (BPMN)

is considered as standard in this ﬁeld (Chinosi and

Trombetta, 2012). It enables a graphical notation

of business processes by mapping the procedure it-

self and involved process entities (e.g., documents) to

modeling elements. Once modeled, processes are ex-

ecuted according to the model. A widely used ap-

proach to improve process models and executions is

Process Mining. Process mining techniques target the

automatic discovery of information from event logs

which contain one or several process cases (Van der

Aalst et al., 2009). Such techniques are able to iden-

tify, for example, that the execution of process steps

in another order maximizes process success. This in-

sight is then considered for future executions through

redesigning the process model. Since event logs rep-

resent model executions, they contain exclusively in-

formation that was previously modeled.

We distinguish between two types of approaches.

We classify techniques that analyze processes by

known information that is already contained in the

process model as approaches that work with intrinsic

parameters. In contrast, approaches that work with

extrinsic parameters incorporate further data sources

or information not yet included in a model.

There are a lot of different approaches that op-

timize processes using merely intrinsic parameters,

e.g., (Gounaris, 2016), (Polyvyanyy et al., 2008),

(Reichert et al., 2012) (Schonenberg et al., 2008).

Also approaches that consider multiple process per-

spectives to improve process analyzability have to be

mentioned in the context of intrinsic techniques, e.g.,

(Front et al., 2017). According to (Radesch

utz et al.,

2008), most business analysis tools do not consider

all relevant data sources or are restricted to a single

data source. This conﬁrms our observation, that there

are only a few approaches based on extrinsic parame-

ters. In (Niedermann et al., 2010) a (semi-) automated

process optimization approach is suggested, which in-

tegrates process and operational data, as well as any

other required data source, e.g., process participant

speciﬁc data. Furthermore, we classify approaches

that add information that cannot be expressed by the

process modeling language as extrinsic. The authors

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

528

of (Wiedmann, 2017) propose the BPMN extension

BPM(N)

Easy

to attach media annotations to model-

ing elements. This provides users additional data

sources that contain information exceeding the ex-

pressive power of a BPMN model.

3 OVERALL CONCEPT

In previous work, we proposed an extrinsic

parameter-based concept to enhance existing

business process models by analyzing image data of

process executions (Fichtner et al., 2020), (Fichtner

et al., 2021). The overall concept consists of ﬁve

steps which are brieﬂy described in this section

while the remainder of this paper is dedicated to the

implementation of the Image Analysis step.

It is assumed that there exists a business process

model for a given process. In the Preparation step,

process tasks are identiﬁed in the model that have be

executed manually by process participants. Each of

these tasks will be analyzed successively for relevant

process details in the overall procedure. In the Deﬁ-

nition step, the process model is modiﬁed regarding

the considered task, such that a video of the task ex-

ecution is recorded or a picture of the initial situation

of the task is taken. The process model is executed

and image data of the task is recorded. After each

execution, the data is labelled according to process

success (Execution and Labelling). The image data

is analyzed by using image mining techniques in the

Image Analysis step. The goal is to identify the re-

gions in all image data that are related to process suc-

cess. The output of this step is the relevant process de-

tail that contains process success related information

and is missing in the process model. To consider the

analyzed process detail in future process executions,

the existing process model is modiﬁed. The modiﬁca-

tion can either be structure-related, text-related or the

detail can be integrated as media annotation. In the

Validation step, it is evaluated whether the preceding

process model modiﬁcation increases process success

or if another task might be responsible for the reduced

process success and has to be considered.

The objective of the approach is to ensure over-

all process success by (i) identifying tasks in a pro-

cess model that were insufﬁciently modeled and (ii)

extending the model by missing but relevant process

details. In contrast to classical quality assurance ap-

proaches, e.g., (Pryk

ari et al., 2010), that evaluate

the direct correlation between a task execution and

its output, the presented concept considers the overall

process success. This allows to identify whether pro-

cess failures originate from single tasks, even though

the execution itself seems to be correct. To realize

the Image Analysis step, we identiﬁed the following

requirements regarding its output:

First, the analyzed process details should not only

contain information extracted from a single positive

example. Although this may be sufﬁcient for im-

provement, the action scope of process participants

will be severely limited. This can be illustrated by an

example: Consider a task in a process model with the

instruction to place a single object on a pallet. Since

the object position is not explicitly stated, the process

participant place the object anywhere on the pallet.

We assume that there are images showing this scene

and that they are labelled according to overall process

success. We further assume that task success depends

on the position of the object on the pallet, i.e., the

task is successful if the object is placed in a certain

region. One single positive example contains the in-

formation that placing an object at a certain position

leads to success. This is a valuable result, however,

restricts the process participants too much and is im-

practical in real process environments. In addition,

the proposed improvement strongly depends on the

selected example. Single examples always include the

risk of being non-representative. Approaches that are

able to consider the full solution space, i.e., the valid

region to place the object, are needed. The idea of

our approach is to take all positive evaluations and to

generalize them. Thus, out of a concrete placement

information from one positive scenario, a region for

placing the object is determined by collecting all pos-

itive examples. Placing the object anywhere in this

region improves the process outcome. We formulate

this procedure as ﬁnding a general conclusion.

Second, the analyzed information must be inter-

pretable for process participants. In the example

above, an analysis of all object positions in positive

labelled images results in a list of coordinates. An in-

struction that contains this list is hard to interpret and

can only hardly be followed. In contrast, a visualiza-

tion where the valid placement region is highlighted

is easier to understand. Therefore the analysis results

have to be transformed in an adequate representation.

4 EXTRACTING RELEVANT

PROCESS DETAILS

In this section we present an implementation for the

Image Analysis step from our overall concept. We fo-

cus on tasks where input speciﬁcations are not mod-

eled prescriptively enough. We identify features that

are relevant for the success of process executions. We

aim at ﬁnding regions of successful process execu-

How LIME Explanation Models Can Be Used to Extend Business Process Models by Relevant Process Details

529

tions in the feature space and at extracting criteria that

separate them from unsuccessful ones. From these

criteria we derive relevant process details which are

integrated into a business process model. Separating a

feature space in multiple classes is the well researched

issue of classiﬁcation. If the feature space is known,

classical machine learning approaches can be used to

determine the class boundaries. If the features have to

be extracted automatically, deep learning mechanisms

are used (Popescu and Lucian, 2014). In the context

of images, convolutional neural networks (CNN) have

proven to be a successful technique. CNNs are used

for prediction, while in most cases the focus is on the

ﬁnal result of a prediction. However, knowing the rea-

son for a prediction and making CNNs explainable

and interpretable is important. Although the parame-

ters that are connected to the decision of a CNN are

difﬁcult to interpret, there are some explanation tech-

niques summarized in (Burkart and Huber, 2021). By

using such an approach, we are able to (i) identify

which features in images are related to successful pro-

cess outcomes and (ii) which values of these features

are required to guarantee process success.

In our implementation approach we use local in-

terpretable model-agnostic explanations (LIME) pro-

vided by (Ribeiro et al., 2016). The concept of LIME

enables the explanation of predictions of any classi-

ﬁer in an interpretable and faithful manner (Ribeiro

et al., 2016). An interpretable model locally around

the prediction is learned and identiﬁed. The authors

of LIME propose a method to explain models through

representative individual predictions. The develop-

ment of LIME was motivated by problem statements

related to trust in the context of system decisions. It

is important to understand the reasons of a decision

and to recognize wrong ones in order to avoid mis-

takes (Dzindolet et al., 2003). LIME complements

existing systems allowing users to assess trust even

when a prediction seems to be correct but is made

for the wrong reasons. Non-experts are enabled to

identify irregularities when explanations are present.

These aspects inspired us to use LIME in the context

of Business Process Management. Knowing the rea-

sons behind a prediction can improve business pro-

cesses fundamentally instead of providing temporary

and case-dependent suggestions for improvement. To

provide an interpretable representation, LIME uses

binary vectors indicating the presence or absence of

a contiguous patch of similar pixels. The explana-

tions are visualized by highlighting decision-relevant

parts in the original images. The authors of LIME

published promising experiments and the source code

of their research

what supports the use of LIME.

https://github.com/marcotcr/lime-experiments

Since LIME only provides local explanations, we an-

alyze the resulting images to derive a global explana-

tion. Giving a global understanding of image expla-

nations is an open research problem (Ribeiro et al.,

2016). We present an idea to tackle this issue for our

experiments.

Furthermore, we address the question of (iii) how

the process model can be extended by analyzed in-

formation. We give an example by modifying a pro-

cess model designed with BPMN.io

. This is an es-

tablished toolkit to view and model BPMN 2.0 dia-

grams. Modeled diagrams can easily be imported and

exported via XML ﬁles. To enrich an existing model

with details, we modify this ﬁle.

4.1 Implementation

Our implementation consists of three parts: First, a

CNN is trained with labelled image data and LIME is

used to explain the classiﬁcation model. The outputs

of the explanation step are images highlighting pros

for the prediction. Second, these images are analyzed

regarding different features and a general conclusion

is derived. Third, analysis results are integrated into a

business process model.

To realize the ﬁrst part, we follow an open source

implementation presenting the usage of LIME

. We

adopt the basic architecture of the CNN which is suf-

ﬁcient for the complexity of our experimental images.

We adapt the parameters for the expected image size

to our data and reduce the value of the batch size and

epochs for system compatibility reasons.

In the second part, we further analyze the results

after using LIME on positive labelled images to derive

a general conclusion. The LIME results are copies of

the original images but contain only those regions that

explain the decision to the positive class. Irrelevant

parts are colored black. The experiments in the next

section show that the remaining image content is in

most cases an object that is involved in the task. We

deﬁne this object as contiguous set of pixels that have

similar colors but that do not have the background

color. A global explanation can be computed based

on local explanations by ﬁnding a possibility to com-

pare super-pixels in different images (Ribeiro et al.,

2016). In our experiments, we address this issue by

analyzing the object visible in each LIME result with

respect to a set of features (color, shape, size and po-

sition (centroid)). We deﬁne this step as ﬁnding a

(last accessed 16 Dec 2021)

https://bpmn.io/ (last accessed 15 Dec 2021)

https://github.com/marcellusruben/All things medium/

blob/main/Lime/LIME image class.ipynb

(last accessed 16 Dec 2021)

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

530

Figure 1: The metal injection molding or extrusion process.

general conclusion. For this purpose, the analysis re-

sults are summarized in feature speciﬁc sets. There is

one set for each feature containing all possible values,

e.g., a color set contains all object colors that were

analyzed across all LIME results. All features and

their corresponding values are relevant process details

which have to be considered during task execution to

enhance process success.

In the third part, the analyzed process details are

integrated in an existing process model. As stated in

Section 3, textual modiﬁcation, structural modiﬁca-

tion or extension by media annotations are possible

ways to enhance a task instruction. For textual mod-

iﬁcations, the model is exported as XML ﬁle and the

value of the appropriate task attribute is extended by

the process detail. For the structural modiﬁcation, ad-

ditional tags have to be created. In case of media an-

notations, an image that represents the process detail

is created and attached to the task.

4.2 Experiments and Results

For our experiments, we design a simple process

model representing a part of the metal injection mold-

ing or extrusion process. These processes are similar,

with the latter replacing molding by extrusion. The

left image in Figure 1 shows this process designed

with BPMN.io. The picture on the right illustrates

the task of manually placing parts on a charging plate

in a real process environment

To evaluate the applicability of LIME in the con-

text of Business Process Management, we recorded

image data showing this placement task. For a ﬁrst

proof of concept, we restrict to simple object shapes

instead of taking parts from industrial environments.

We prepare our images by using object recognition

techniques. This excludes disturbing factors regard-

ing image quality, such as noise or uncontrollable

lighting conditions. We recorded 1000 images per

https://www.youtube.com/watch?v=QaMdjKE7vT8

(last accessed 17 Dec 2021)

experiment to ensure a sufﬁciently large database to

compute a representative result. The classiﬁcation

problem is limited to two classes, resulting in 500

negative examples (unsuccessful process executions)

and 500 positive examples (successful process execu-

tions). Each image shows an input constellation of the

task. We know the criteria that distinguishes positive

and negative labelled image data. In real applications

these criteria are not known but should be analyzed in

the Image Analysis step. The analyzed criteria cor-

respond to our deﬁnition of relevant process details.

However, we exploit the knowledge of the criteria in

our experiments to validate the classiﬁcation results.

An explanation of all LIME parameters can be

found in the ofﬁcial code documentation. We adopt

the default settings for most of them except two pa-

rameters: We restrict the number of labels for which

an explanation is made to two classes (top labels =

2). The parameter (num f eatures) is adjusted for

each experiment. It deﬁnes the number of similar

pixel regions to be included in the explanation.

Experiment 1: Color

In a ﬁrst experiment, we use a simple scenario where

the color of an object is the decisive criterion. Each

image shows one rectangular object on a plane while

images with blue objects are labelled positive and im-

ages with green ones are labelled negative. We show

examples of both classes including their processed

variants after using object recognition techniques in

Figure 2. The object positions are determined ran-

domly and, just as the shape, should not be a relevant

feature. While the decision criteria is obvious for a

human being, a CNN has to be trained in order to rec-

ognize that only one characteristic is important. In

contrast, we will see later that the system is able to ex-

tract relevant features in more complex scenes. This

is an important aspect when it comes to real process

environments. For this experiment we use 800 im-

ages for the training and 200 for the validation of the

CNN. Based on the positive labelled images (cf. Fig-

ure 2b), local explanations are computed using LIME.

Examples of results with num f eatures = 4 can be

seen in Figure 3. The LIME results show that the

CNN’s decision strongly depends on the object and its

(a) Negative example. (b) Positive example.

Figure 2: Image data for experiment 1. Per example: origi-

nal images (left) and their processed images (right).

How LIME Explanation Models Can Be Used to Extend Business Process Models by Relevant Process Details

531

Figure 3: LIME results of experiment 1.

characteristics, while the background has no impact.

The other computed results are comparable to those

shown here. In 22% of all cases, the images contain

only the object, while the rest of the image is black-

ened (cf. left images in Figure 3). In the remaining

cases, some small parts of the background are visi-

ble (cf. right images in Figure 3). To derive a general

conclusion, the characteristics of the remaining image

content, i.e., the object, are analyzed. Pixels that do

not have the background color are considered. There-

fore, images where small parts of the background can

be seen are processed as well. The analysis results

in a list of valid positions and sizes as well as single

values for the features shape and color. If these val-

ues are considered during task execution, the process

is successful. For this purpose, the business process

model is extended by this information. Therefore we

edit the XML ﬁle describing the process model and

change the name of the task in the respective line. We

show an excerpt of the modiﬁed ﬁle in Listing 1.

<bpmn:task id=”A1” name=”place parts on charging plate;

Position:{(51, 159), ...};

Size:{(44, 116), ...};

Shape:{’rectangle’};

Color:{’blue’}”>...</bpmn:task>

Listing 1: Modiﬁcation of the task-related tag.

To support readability, this information can also be

appended as text annotation (cf. Figure 4). How-

ever, this information does not only contain the rel-

evant detail, i.e., the blue color. Furthermore, the po-

sition and size lists are given which are hard to in-

terpret. At this point, the analyzed information has

to be related to the process context. In our experi-

ment, all involved objects are rectangular and of same

size. So the values for the features shape and size may

be neglected. Considering the position list and object

size, it can be derived that all positions on the plane

are valid. Therefore also the position is no decisive

feature. However, all objects are either blue or green

conﬁrming the color as relevant feature. The ﬁnal re-

Figure 4: Analyzed details added as text annotation.

sult can be interpreted as instruction to place only blue

parts on the charging plate.

Experiment 2: Position and Shape

In a more complex placing experiment, each image

shows four objects. Two of them are rectangular and

the other two are circular. The color of all objects is

the same. The object positions are determined ran-

domly. We exclude intersections and ensure that no

object protrudes beyond the image border. An im-

age is labelled as positive, if at least one circular ob-

ject is positioned in the upper seventh of the scene.

Thus, the features shape and position are the relevant

ones. Figure 5 shows examples leading to unsuccess-

ful and successful process executions. Due to space

limitations we show only the resulting images after

using object recognition techniques. Out of the total

1000 images, 800 are used for training of the CNN

and 200 are used for validation. Results after apply-

ing LIME with parameter num f eatures = 2 are pre-

sented in Figure 6a. The left image shows the result

after applying LIME on the positive labelled image

presented in the left of Figure 5b. In all resulting im-

ages, one circular object is highlighted in the upper

part of the plane. Among them, 29% of the results are

comparable to the left image of Figure 6a. The others

are comparable to the right image and either do not

contain the circular object completely or additionally

show small areas of the background. However, none

of the results contain a rectangular object or both cir-

cular objects. This is an ideal result since all LIME

images represent the condition that one circular object

has to be placed in the upper part of the scene. Besides

a list of positions and sizes, the analysis step outputs

”red” as single color value and ”circle, pentagon” as

shape values. The wrong shape ”pentagon” occurs in

few cases. It results from LIME explanations where

the circular object is not completely visible (cf. right

image in Figure 6a). Since the position list is hard to

interpret, we suggest another representation. We sug-

gest to create an image that shows the region of valid

positions for placing an object (cf. Figure 6b). This

representation allows to efﬁciently get an overview of

all placement options. The region is computed by

ﬁnding the minimal bounding box that encloses all

object position values that are analyzed from LIME

(a) Negative examples. (b) Positive examples.

Figure 5: Image data for experiment 2.

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

532

(a) The resulting images highlight

one circular object.

(b) Valid region

to place objects.

Figure 6: LIME results of positive samples from experiment

2 (left). Bounding box of all valid object positions (right).

results. In this scenario, we suggest to attach this

image to the respective task in the business process

model by using media annotation concepts proposed

by (Wiedmann, 2017). Process participants can view

the attached image during process execution to get a

visual extension of the task instruction. Furthermore,

the task instruction itself is adjusted to include the an-

alyzed and relevant features (color, shape). The XML

ﬁle of the process model is modiﬁed according to ex-

periment 1.

4.3 Evaluation and Discussion

The experiments conﬁrm that our approach is able to

identify relevant process details from images of pro-

cess tasks. The relevant feature space can be deter-

mined by using CNNs without making any further

assumptions about features. From a Business Pro-

cess Management perspective, this is an important as-

pect. Since relevant details and features are usually

not known in advance (cf. Section 1), assumptions

regarding the feature space cannot be made. The con-

cept of LIME is able to identify and visualize regions

in the computed feature space that are relevant for

successful process execution. In our experiments, the

images created by LIME highlight single objects that

take part in the scene. Considering the criterion that

distinguishes negative from positive labelled images

per use case, LIME produces expected and correct re-

sults in both experiments. The LIME results explain

that the relevant information regarding process suc-

cess is somehow related to the highlighted object. The

rest of the image is irrelevant for the decision of the

CNN. In the second experiment, pre-ﬁltering to rel-

evant content is essential to derive a general conclu-

sion. It indicates that the rectangular objects are not

relevant for process success. It further explains that

only one circular object is relevant for the CNN’s de-

cision for the positive class. By analyzing the features

of this remaining object across all LIME results, we

are able to compute a general conclusion. Finding

such a conclusion meets an important in the context

of Business Process Management. As explained at the

end of Section 3, a local explanation based on a sin-

gle positive example is not sufﬁcient as it restricts the

scope of action of a process participant unnecessarily

far.

In the experiments, we demonstrate the applica-

bility of the concept in simple scenarios and with pre-

processed image data using object recognition tech-

niques. In real process environments, scenes can be-

come much more complex. Although the presented

approach also works with more complex input data,

we propose to optimize it for more complex scenes.

To meet associated requirements, the presented image

analysis technique should be interchanged with more

powerful image analysis techniques. Complex images

require robust techniques that are able to explore large

feature spaces. Furthermore, derived general conclu-

sions have to be optimized in case of more complex

scenes. To automate this issue the implementation has

to be extended as follows. The general conclusion

contains lists of object features and a set of values per

feature. Each set has to be checked whether it con-

tains all possible values that can occur in the process

task. If this is true, the related object feature is not

a relevant detail. For example, in experiment 1 the

object shape is not a relevant feature since all objects

involved in the process task are rectangular. Finally, a

sufﬁcient number of image data for training and val-

idating the CNN may not be available in real process

environments. In the case of limited data, we propose

to integrate data augmentation techniques into the im-

plementation.

5 CONCLUSION AND FUTURE

WORK

In this work, we present an approach to identify miss-

ing process details of business process models that are

relevant for an overall process success. In the pro-

posed implementation, a CNN is trained with image

data showing task scenes. The images are recorded

during task execution and are labelled according to

overall process success. We use LIME to explain the

prediction of the CNN for samples of successful pro-

cess executions. The results are images containing

prediction-relevant regions. Across all results, these

regions are analyzed using image analysis techniques

to derive a general conclusion. It contains relevant

features and values to achieve process success. This

information is integrated into the process model. We

evaluate our method with experiments using image

data showing simpliﬁed scenarios from manufactur-

ing process environments. Our experiments conﬁrm

How LIME Explanation Models Can Be Used to Extend Business Process Models by Relevant Process Details

533

that LIME and image mining techniques can be used

to improve business processes. The results show that

our approach is able to identify relevant process de-

tails. The process of metal injection molding or ex-

trusion is taken as an example from real-world envi-

ronments. The presented concept is applicable for a

large set of similar processes that involve process par-

ticipants in pick-and-place tasks. However, its ﬂexi-

bility allows to extend its scope of application to other

tasks.

Future work should focus on using more power-

ful image mining techniques that increase the robust-

ness and accuracy of the Image Analysis step. We

aim to extend our experiments with more complex

image data from real process environments. We fur-

ther plan to evaluate the proposed ways of integrating

process details in existing process models in a user

study. This includes the investigation of how a certain

feature must be represented in order to be interpreted

correctly and efﬁciently by process participants.

REFERENCES

Van der Aalst, W. M. P., Van Dongen, B. F., G

unther, C. W.,

Rozinat, A., Verbeek, E., and Weijters, T. (2009).

Prom: The process mining toolkit, in bpm (demos).

In CEUR Workshop Proceedings, volume 489.

Becker, J., Rosemann, M., and von Uthmann, C. (2000).

Guidelines of Business Process Modeling, pages 30–

49. Springer Berlin Heidelberg, Berlin, Heidelberg.

Burkart, N. and Huber, M. F. (2021). A survey on the ex-

plainability of supervised machine learning. Journal

of Artiﬁcial Intelligence Research, 70:245–317.

Chinosi, M. and Trombetta, A. (2012). Bpmn: An intro-

duction to the standard. Computer Standards & Inter-

faces, 34(1):124–134.

Dzindolet, M. T., Peterson, S. A., Pomranky, R. A., Pierce,

L. G., G, L., and Beck, H. P. (2003). The role of

trust in automation reliance. International journal of

human-computer studies, 58(6):697–718.

Fichtner, M., Sch

onig, S., and Jablonski, S. (2020). Pro-

cess management enhancement by using image min-

ing techniques: A position paper. In Proceedings of

the 22nd International Conference on Enterprise In-

formation Systems, volume 1 of ICEIS, pages 249–

255.

Fichtner, M., Sch

onig, S., and Jablonski, S. (2021). Us-

ing image mining techniques from a business process

perspective. In Enterprise Information Systems, pages

62–83, Cham. Springer International Publishing.

Front, A., Rieu, D., Santorum, M., and Movahedian, F.

(2017). A participative end-user method for multi-

perspective business process elicitation and improve-

ment. Software & Systems Modeling, 16(3):691–714.

Gounaris, A. (2016). Towards automated performance op-

timization of bpmn business processes. In East Euro-

pean Conference on Advances in Databases and In-

formation Systems, pages 19–28. Springer.

Kluza, K., Baran, M., Bobek, S., and Nalepa, G. J. (2013).

Overview of recommendation techniques in business

process modeling. In Proceedings of 9th Workshop

on Knowledge Engineering and Software Engineering

(KESE9), pages 46–57. Citeseer.

Mendling, J., Reijers, H. A., and van der Aalst, W. M. P.

(2010). Seven process modeling guidelines (7pmg).

Information and Software Technology, 52(2):127–

136.

Niedermann, F., Radesch

utz, S., and Mitschang, B. (2010).

Deep business optimization: A platform for auto-

mated process optimization. INFORMATIK 2010–

Business Process and Service Science–Proceedings of

ISSS and BPSC.

Polyvyanyy, A., Smirnov, S., and Weske, M. (2008). Pro-

cess model abstraction: A slider approach. In 2008

12th International IEEE Enterprise Distributed Ob-

ject Computing Conference, pages 325–331. IEEE.

Popescu, M. C. and Lucian, M. S. (2014). Feature extrac-

tion, feature selection and machine learning for im-

age classiﬁcation: A case study. In 2014 International

Conference on Optimization of Electrical and Elec-

tronic Equipment (OPTIM), pages 968–973. IEEE.

Pryk

ari, T., Czajkowsk, J., Alarousu, E., and Myllyl

a, R.

(2010). Optical coherence tomography as an accurate

inspection and quality evaluation technique in paper

industry. Optical review, 17(3):218–222.

Radesch

utz, S., Mitschang, B., and Leymann, F. (2008).

Matching of process data and operational data for a

deep business analysis. In Enterprise Interoperability

III, pages 171–182. Springer.

Reichert, M., Kolb, J., Bobrik, R., and Bauer, T. (2012).

Enabling personalized visualization of large business

processes through parameterizable views. In Proceed-

ings of the 27th Annual ACM Symposium on Applied

Computing, pages 1653–1660.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ” why

should i trust you?” explaining the predictions of any

classiﬁer. In Proceedings of the 22nd ACM SIGKDD

international conference on knowledge discovery and

data mining, pages 1135–1144.

Schmidt, R., M

ohring, M., Zimmermann, A., H

arting, R.-

C., and Keller, B. (2016). Potentials of image mining

for business process management. In Intelligent Deci-

sion Technologies 2016, pages 429–440. Springer.

Schonenberg, H., Weber, B., van Dongen, B., and van der

Aalst, W. M. P. (2008). Supporting ﬂexible processes

through recommendations based on history. In In-

ternational Conference on Business Process Manage-

ment, pages 51–66. Springer.

Wiedmann, P. C. K. (2017). Agiles

Gesch

aftsprozessmanagement auf Basis ge-

brauchssprachlicher Modellierung. PhD thesis,

University of Bayreuth, Germany.

ICEIS 2022 - 24th International Conference on Enterprise Information Systems

534