Behavioral Recommender System for Process Automation Steps

Mohammadreza Fani Sani

1 a

, Fatemeh Nikraftar

, Michal Sroka

1 b

and Andrea Burattin

2 c

Microsoft Development Copenhagen, Denmark

Technical University of Denmark, Copenhagen, Denmark

Keywords:

Process Mining, Process Automation, Microsoft Process Automation, Recommender Systems.

Abstract:

Process automation is used to increase the performance of processes. One of the leading process automa-

tion tools is Microsoft Process Advisor. This tool requires users to select the corresponding connectors for

the automation of different tasks, which can be a challenging endeavor for users who have limited business

knowledge as there are various connectors and templates exist. To overcome this challenge, we present a

process-aware recommender system for connectors that eases the labeling task for end users. The results of

applying this method to real event logs indicate that it can recommend relevant connectors and, therefore, the

usage of the same mechanism might be generalized to broader contexts.

1 INTRODUCTION

Process mining bridges data science and business

process management. It consists of process discov-

ery, conformance checking, and process enhance-

ment. Process discovery generates a process model

from event logs, conformance checking compares the

logs and model, and process enhancement provides

insights for improvement (van der Aalst, 2016).

Gartner estimated that process enhancement will

get more attention in the upcoming years

. In

this sub-ﬁeld, one of the research directions that

helps process improvement is robotic process automa-

tion (Leno et al., 2020). The goal of this ﬁeld is to de-

tect the bottlenecks of the process and try to automate

the tasks that require more time. In this regard, we

aim to detect and handle routine and administrative

tasks automatically using current technologies. As

a positive effect, we will reduce the resources’ costs

and the required times to handle the process instances.

However, robotic process automation techniques also

have their limits, and one of the fundamental chal-

lenges is selecting a suitable activity or process to au-

tomate, thus understanding which components should

be involved in the automation. This task, in the litera-

ture, is referred to as task mining (Syed et al., 2020).

https://orcid.org/0000-0003-3152-2103

https://orcid.org/0000-0002-7505-2521

https://orcid.org/0000-0002-0837-0183

See https://www.gartner.com/doc/reprints?id=1-28S

A9BAA&ct=220118&st=sb

One of the leading process automation tools is

Microsoft Process Advisor (MPA)

. This tool pro-

vides the opportunity to automate manual and repeti-

tive tasks that often take a lot of time. For example,

in an accounting company, a resource should ﬁnd the

amount of a receipt and convert its currency by ﬁnd-

ing the rates in a browser and ﬁnally send the con-

verted fee to the customer by email. Using MPA, it is

possible to record the event log based on the execu-

tion of some process instances and provide the corre-

sponding process automation steps. To provide such

automation, the user needs to connect tasks to the

corresponding connectors, i.e., proxies or wrappers

around an API that lets the underlying service com-

municate with several Microsoft products. Currently,

MPA supports more than 275 connectors and thou-

sands of pre-built templates that allow for easy inte-

gration of popular end-services in workﬂow develop-

ment/improvement. One of the challenges of automa-

tion in MPA is selecting the most suitable connector

after recording the process. Without having knowl-

edge of the process, the selection of proper connectors

could be a challenging task.

To overcome this challenge, in this paper, we pro-

pose a recommender system that provides the best

connectors for automating a process. To evaluate its

accuracy, we develop it and apply it to 50 real process

scenarios (Sroka and Fani Sani, 2022). The results in-

dicate that the proposed recommender system is able

See https://powerautomate.microsoft.com/en-us/proc

ess-advisor/

Fani Sani, M., Nikraftar, F., Sroka, M. and Burattin, A.

Behavioral Recommender System for Process Automation Steps.

DOI: 10.5220/0012060800003541

In Proceedings of the 12th International Conference on Data Science, Technology and Applications (DATA 2023), pages 255-262

ISBN: 978-989-758-664-4; ISSN: 2184-285X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

255

to help the end-user select the corresponding connec-

tor.

The structure of the remaining part of the paper is

as follows. First, we provide related work in the area

of process mining and recommender systems. There-

after, the preliminaries that ease the rest of the paper.

After that, in Section 4, we explain the proposed rec-

ommender system. Section 5 provides the results of

evaluating the proposed method. Finally, in Section

6, we conclude the paper and propose some new di-

rections to continue this research.

2 RELATED WORK

Several works have been done in the areas of rec-

ommender systems and process automation in pro-

cess mining and a good overview of these is available

in (Eili et al., 2021), where different systems are sur-

veyed.

Different process discovery techniques exist, each

employing various parameters that sometimes make it

difﬁcult which technique and setting should be used.

There are some works, such as (Ribeiro et al., 2014;

Wang et al., 2012), that aim to help users to discover

more suitable process models. In (Jr. et al., 2021),

the authors use a meta-learning approach that is able

to recommend a suitable process discovery conﬁgu-

ration with high accuracy. Moreover, the work pre-

sented in (Schonenberg et al., 2008) discusses the

possibility of applying process mining techniques to

recommend to end-users what should be done in the

next phase of the process. Furthermore, (Seeliger

et al., 2018) proposes an interactive recommender

system that provides suggestions for visualising the

process model. In (Terragni and Hassani, 2018), the

authors propose some possibilities to use process min-

ing on the logs of users’ interactions on websites to

explore the customer journey, predict their future ac-

tivities, and recommend actions that maximize partic-

ular performance indicators.

In (Geyer-Klingeberg et al., 2018), the authors

present a process mining technique to enable effective

RPA activities towards process improvement. More-

over, (Wanner et al., 2019) discusses the capabilities

of processes-based techniques with RPA and proposes

an automatable indicator system as well as RPA activ-

ities to maximize the automation investment.

In (Mayr et al., 2022), a survey on task mining

has been reported and several applications and chal-

lenges in front of this research are represented. More-

over, (Choi et al., 2022) has proposed a tool to record

the interactions with user interfaces and to generate an

event log that can be used to bridge the gap between

process mining and RPA by detecting the tasks that

can be automated.

Some tools like MPA

provides tem-

plates/connectors to automate tasks and activities.

We aim to provide a connector recommender system

to help users in the selection of corresponding

connectors.

3 PRELIMINARIES

In this section, some process mining concepts are dis-

cussed. In process mining, we use events to pro-

vide insights into the execution of business processes.

Each event is related to speciﬁc activities of the under-

lying process. Furthermore, we refer to a collection of

events related to a speciﬁc process instance as a case.

Both cases and events may have different attributes.

An event log that is a collection of events and cases is

deﬁned as follows.

Deﬁnition 3 (Event Log). Let E be the universe

of events, C be the universe of cases, AT be the

universe of attributes, and U be the universe of at-

tribute values. Moreover, let C ⊆ C be a non-empty

set of cases, let E ⊆ E be a non-empty set of events,

and let AT ⊆ AT be a set of attributes. We deﬁne

(C, E, π

, π

) as an event log, where π

: C × AT 7→

U and π

: E × AT 7→ U. Any event in the event

log has a case, therefore, ̸ ∃

e∈E

(π

(e, case) ̸∈ C) and

e∈E

(π

(e, case))=C.

Furthermore, let A ⊆ U be the universe of activ-

ities and let A

∗

be the universe of sequences of ac-

tivities. For any e ∈ E, function π

(e, activity) ∈ A,

which means that any event in the event log has an ac-

tivity. Moreover, for any c ∈C function π

(c,trace) ∈

∗

\ {⟨⟩} that means any case in the event log has

a trace (i.e., a sequence of events). Having a trace

of a case and the activities of the events in the trace,

we can have a variant of a case, i.e., π

(c, variant) ∈

∗

\ {⟨⟩}.

Therefore, mandatory attributes for events are

case and activity, and for cases are trace and variant.

There are different notations to show a process

model, e.g., Petri net (Petri and Reisig, 2008) and

BPMN (Chinosi and Trombetta, 2012) models. How-

ever, it is also possible to describe a process model by

the complete sequence of activities that is possible to

execute by it. In the following, we deﬁne a process

model and a process discovery algorithm.

Deﬁnition 4 (Process Model and Process Discov-

ery). Let A be the universe of activities. M =

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

256

P(A

∗

) \ {} is the universe of process models where

P(X) is the powerset of set X.

Moreover, let E L be the universe of event logs,

pd : E L → M is a process discovery algorithm that

returns a process model for each event log.

To compute how a process model and an event log

conform to each other, we use conformance checking

measures. There are different conformance check-

ing measures that are reported in the literature, e.g.,

alignments (Adriansyah and Buijs, 2012) and token-

replay (Rozinat and van der Aalst, 2008). In the fol-

lowing, conformance checking is deﬁned.

Deﬁnition 5 (Conformance Checking). Let A be

the universe of activities and let M denote the uni-

verse of process models. cc:A

∗

×M → [0, 1] is a con-

formance checking function that returns how much

behavior in a sequence of activities is represented in a

process model.

The conformance of a process model and an event

log is the average conformance value of variants of

all its traces (π

(c, variant)). However, in this paper,

we use the conformance checking at the variant level

(i.e., we do not count repetitions of the same trace).

The set of all variants in an event log (C, E, π

, π

) is

{π

(c, variant)|c ∈ C}.

As explained, traces and variants are the se-

quences (of events and activities). We are able

to use projection function on the sequence σ =

⟨x

, x

, .. ., x

⟩. So, if Q ⊆ X,↿

: X

∗

→ Q

∗

is a pro-

jection function that returns the concatenation of ele-

ments of the input sequence, i.e., σ↿

= ⟨x

↾

. .. x

↿

. .. x

↿

⟩. For instance, ⟨a, b,b, c,d, e, f , h⟩↿

{a,b,h}

⟨a, b, b, h⟩. Note that, x↿

= ⟨⟩ if x ̸∈ Q.

4 PROCESS AWARE

RECOMMENDER SYSTEM

In this section, we explained the proposed process-

aware connector recommender system. We ﬁrst need

to deﬁne a connector recommender system.

Deﬁnition 6 (Connector Recommender). Let E

denote the universe of activities and let M be the uni-

verse of process models. Moreover, let L be the uni-

verse of connectors. We deﬁne cr : A

∗

× P(M ) →

P(L) as a connector recommender that receives a

trace (i.e., a sequence of activities) and a set of pro-

cess models, and returns a set of labels. Finally, cr

is a speciﬁc type of connector recommender that for

trace t ∈ A

∗

and process models M

⊆ M ∧|M

| ≥ k,

we have |cr

(E, M

)| = k.

In other words, we recommend the connectors for

the variants in the event log (can be extracted by

(c, variant)). The schematic view of this method is

presented in Figure 1. The proposed method consists

of two phases, i.e., Train and Application. In the train-

ing phase, as the proposed approach is a supervised

method, we ﬁrst need to provide the training/labeled

event log(s). In other words, we need to have some

traces labeled by their connectors.

It should be noted that one case or its trace can

have one or more than one connectors. Therefore,

a trace will be divided into some subsequences and

each subsequence can relate to one connector. In the

following, we deﬁne the labeled event log.

Deﬁnition 1 (Labeling Event Log). Let E be the

universe of events and let L ⊂ U be the universe of

labels. We deﬁne π

(e, label) ∈ L ∪ τ that assigns a

label or no value to each event.

For each trace σ in the labeled event log, we may

have 0 to |σ| labels. It should be noted that in re-

ality, we usually assign labels to subsequences of a

trace and we assign a label to all the events of a subse-

quence. Moreover, note that an activity can be labeled

with different connectors in different events.

Afterward, we project the event log based on their

assigned connectors, consequently, we will have one

sub-log for each connector. We can use the following

function for this purpose.

Deﬁnition 2 (Event Log Projection). Let E L be

the universe of event logs and let EL=(C, E, π

, π

)

∈ E L be a labeled event log. Moreover, let L ∈ L be

a set of labels. We deﬁne ep : E L × L 7→ P(E L)

as an event log projection function that receives a

labeled event log and a set of connectors and re-

turn a set of projected event logs where ep(EL, L) =

{(C

, E

, π

)|l ∈ L ∧



= {e ∈ E|π

(e, label) =

l} ∧C

= {c ∈ C||π

(c,trace)↿ E

| > 0}}



In other words, for each connector, we gather all

the events (and the cases if it remains at least one

event in it) with the corresponding label in a sub-

log. Note that

,π

)∈ep(EL,L)

) = {}; how-

ever,

,π

)∈ep(EL,L)

) ≤ |E| as it is possible

that some of the events have no labels.

Thereafter, for each of the projected sub-logs, we

discover a process model that represents the general

behavior in that sub-log. It is not necessary to use

similar parameters for different sub-logs. Note that as

we deﬁne a process model as a possible sequence of

activities, we can also consider the variants of the pro-

jected event logs (i.e., M

={π

(c, variant)|c ∈C

}) as

Behavioral Recommender System for Process Automation Steps

257

Figure 1: A schematic view of the proposed connector recommender system. In the training phase, logs with the labels (i.e.,

the expected connector) are provided and a model is extracted for each label. The recommendation step (i.e., application)

computes the conformance of the given trace against all models and selects the recommender associated with the highest

conformance (i.e., the model that best ﬁts the behavior of the recommender).

Table 1: Some information about the event logs that are used in the evaluation.

Connector Traces Labeled Traces Train Traces Test Traces

Approvals 30 30 23 7

Googlecalendar 30 26 20 6

Microsoftforms 21 21 16 5

Ofﬁce365users 21 18 14 4

Onenote 24 15 11 4

Outlook 21 21 16 5

Planner 21 18 13 5

RSS 18 12 8 4

Sendmail 18 18 13 5

Sharepoint 18 15 11 4

their process models. The discovered process models

will be passed to the next phase. In other words, the

output of the training phase is a set of process models

for the connectors.

In the application phase, we aim to recommend

some connectors for a given trace. For this purpose,

we compute the conformance of the trace with all the

discovered process models. The higher conformance

values mean it is more likely that the trace relates to

the corresponding connector. Thereafter, we recom-

mend the top-k connectors, i.e., the k connectors with

the highest conformance value.

5 EVALUATION

In this section, we evaluate the proposed recom-

mender system by applying it to real event logs. First,

we explain the dataset that was used in the evaluation

and the detail of the implementation and evaluation.

Afterward, the evaluation results are discussed.

5.1 Experimental Setting

To evaluate the proposed connector recommender

system, we used a dataset containing 50 processes,

each of which contains 3 or more process instances

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

258

Table 2: Information about the discovered process models.

Model Places Arcs Transitions Activities Shared Activities Fitness Precision

Approvals 65 212 99 27 136 0.964 0.209

Googlecalendar 36 124 59 20 114 0.996 0.238

Microsoftforms 56 182 85 26 130 0.987 0.123

Ofﬁce365users 48 160 76 19 123 0.960 0.154

Onenote 34 106 50 19 91 0.967 0.252

Outlook 59 196 93 26 128 0.992 0.199

Planner 58 192 92 23 104 0.979 0.124

RSS 27 86 43 13 74 0.987 0.353

Sendmail 49 152 70 19 119 1 0.272

Sharepoint 32 104 49 18 113 1 0.242

(i.e., cases)

. We have labeled the event logs using

experts with business knowledge, which gives us the

training dataset to be used with the proposed method.

There are 25 connectors (i.e., the labels) in the event

log, however, we preprocess the dataset and keep only

the 10 most frequent ones. We split the dataset into

train and test parts: in the train event logs, we had 145

cases (on average 14.5 cases per connector); whereas

in the test event logs, there were 49 cases. Some

statistics of this event log is presented in Table 1

We have implemented the proposed method in

Python, using the PM4Py library (Berti et al., 2019)

To discover process models, we have used the In-

ductive Miner (Leemans et al., 2013) with the noise

threshold value equal to 0.1 (to have more speciﬁc

process models). Furthermore, to compute the confor-

mance value, we have used the alignment technique.

To recommend connectors, we use cr

with different

k-values.

To evaluate the accuracy of the recommendation,

we have used the following formula:

RetrieveRate =

|{LabeledConnector} ∩{RecommendedConnectors}|

|{RecommendedConnectors}|

(1)

The higher value means a more accurate recom-

mendation.

5.2 Experimental Results

As explained, we ﬁrst need to discover process mod-

els using the projected event logs. Some information

The event log is available at https://github.com/micro

soft/50BusinessAssignmentsLog

This preprocessed labeled dataset is available at https:

//github.com/nikraftarf/Recommender-System-Based-on-p

rocessMining/tree/main/Data/Main-Data(Second-round).

The implementation is available at https://bit.ly/3PO

X7z5

about the discovered process models is given in Ta-

ble 2. In this table, activities indicate the number of

labeled transitions in process models, and shared ac-

tivities show how many times the activities appear in

the other process models. The ﬁtness and precision

values indicate the quality of process models. As it

can be noted in the table, for all process models, we

have high ﬁtness and low precision, which is mainly

because of the noise threshold that is used (i.e., 0.1).

Among the models, RSS has the fewest activities and

shared activities. The highest

shared activities

activities

value be-

longs to the process model of Ofﬁce365users connec-

tor.

Fig. 2 shows the discovered process model of the

RSS connector.

In the next step, we used the discovered process

models to detect which connectors should be recom-

mended for each trace in the test event logs. The re-

sults of using the proposed connector recommender

system on the test data are presented in Table 3.

The results show that for some connectors, e.g., Of-

ﬁce365users, the labeled connector exists in the rec-

ommended connectors, even by recommending a few

connectors. It is mainly because the discovered pro-

cess model for these connectors is more distinguish-

able. On the other hand, for RSS connector, even by

increasing the number of recommended connectors,

the labeled connector does not exist in the list of rec-

ommendations. The low accuracy for the recommen-

dations of RSS’s cases could be related to the few gen-

eral labeled activities in its process model. Moreover,

as we can see in Table 2, the precision of the pro-

cess model of this connector is higher compared to

the other process models.

It can be seen that by increasing the number of

recommended connectors, i.e., k, we have a higher

retrieve rate. However, recommending too many

connectors can reduce the beneﬁt of the proposed

method. Therefore, ﬁnding optimal ways of properly

Behavioral Recommender System for Process Automation Steps

259

Figure 2: The process model (in Petri net description) discovered on projected event log of RSS connector.

Table 3: Retrieve Rate of the proposed method using different k values for different connectors.

Retrieve Rate

Connector Records k = 1 k = 2 k = 3 k = 4 k = 5

Approvals 7 0.57 0.71 0.71 1 1

Googlecalendar 6 0.5 0.5 0.67 1 1

Microsoftforms 5 0.6 0.6 0.8 1 1

Ofﬁce365users 4 0.75 0.75 1 1 1

Onenote 4 0.25 0.25 0.5 0.5 0.5

Outlook 5 0.6 0.6 0.8 0.8 0.8

Planner 5 0.6 0.8 0.8 1 1

RSS 4 0 0 0 0 0.25

Sendmail 5 0.2 0.8 0.8 0.8 1

Sharepoint 5 0.4 0.4 0.6 0.8 0.8

Table 4: The number of times that each connector is recommended when different k-values are used.

Approvals Googlecalendar Microsoftforms Ofﬁce365users Onenote Outlook Planner RSS Sendmail Sharepoint

k = 1 3 3 5 6 1 9 9 0 9 2

k = 2 8 5 12 9 2 12 16 0 20 5

k = 3 11 6 24 15 2 14 20 1 25 7

conﬁguring k represents a signiﬁcant future direction

of research.

In Table 4, we have provided the frequency of dif-

ferent connectors when different k-values are used.

Suppose that connector c is recommended as the mth

connector if we recommend k ≥ m connectors, this

connector will be presented among them too. The

results indicate that we recommend some connectors

like Sendmail and Microsoftforms more often. Con-

versely, we infrequently recommend some connectors

like RSS and Onenote which are recommended only 3

times. The reason for this is due to the absolute fre-

quency of the activities in the projected event logs:

some connectors are more common to appear in all

logs hence these are able to replay traces of other con-

nectors with higher ﬁtness value.

5.3 Discussion and Limitation

While the work has been validated with experiments

on real datasets, which revealed the efﬁcacy of the

technique, further investigations are needed to gener-

alize the conclusions.

The proposed method assumes that most activities

in the projected training event logs exist in the traces

belonging to the connector we want to recommend.

To overcome this limitation, NLP techniques such as

semantic similarity methods (Hayes and Henderson,

2021) can be applied to deﬁne equivalence classes for

activities with different names but same action.

The performance of discovered process model is

highly dependent on the quality of discovered pro-

cess models. For example, if we discover a ﬂower

model (van der Aalst et al., 2010), the process model

will have high ﬁtness values for different traces. The

mining algorithm and the used setting for it can be

seen as a hyperparameter of the technique and, like

k, requires further investigation. But, contextual in-

formation is required to properly conﬁgure the tech-

nique.

The proposed method is a supervised technique

and requires labeled event logs. Labeling can be ex-

pensive and time-consuming, but tools can be used to

provide a rough estimate of the label. Additionally,

labeling is only required once, and pre-trained mod-

els (Qiu et al., 2020) can be used to reduce the effort.

Transform learning and pre-trained models (Qiu et al.,

2020) can also be used due to the limited set of labels

(i.e., the set of all connectors in MPA).

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

260

6 CONCLUSION

The proposed process-aware recommender system is

a promising method for suggesting optimal steps to-

wards process automation. By leveraging process

models discovered from sub-logs created based on ac-

tivity projections and assigned connectors in labeled

event logs, the system uses user behavior to make rec-

ommendations. The results from applying this system

to real event logs, previously labeled by an expert,

demonstrate that for most connectors, the system is

able to recommend the expected automation steps ac-

curately.

To further improve the system’s effectiveness, fu-

ture research will consider additional data attributes,

such as the resources that execute activities and time

dimensions, and explore other techniques such as as-

sociation rule mining. Finally, classical data mining

approaches will also be considered to compare the

system’s performance with alternative approaches.

Overall, these efforts will enhance the system’s rec-

ommendation capabilities and further validate its use-

fulness in process automation.

REFERENCES

Adriansyah, A. and Buijs, J. C. A. M. (2012). Mining

process performance from event logs. In Business

Process Management Workshops - BPM 2012 Inter-

national Workshops, Tallinn, Estonia, September 3,

2012. Revised Papers, pages 217–218.

Berti, A., Van Zelst, S. J., and van der Aalst, W. (2019).

Process mining for python (pm4py): bridging the gap

between process-and data science.

Chinosi, M. and Trombetta, A. (2012). BPMN: an intro-

duction to the standard. Comput. Stand. Interfaces,

34(1):124–134.

Choi, D., R’bigui, H., and Cho, C. (2022). Enabling the

gab between RPA and process mining: User interface

interactions recorder. IEEE Access, 10:39604–39612.

Eili, M. Y., Rezaeenour, J., and Sani, M. F. (2021). A

systematic literature review on process-aware recom-

mender systems. CoRR, abs/2103.16654.

Geyer-Klingeberg, J., Nakladal, J., Baldauf, F., and Veit,

F. (2018). Process mining and robotic process au-

tomation: A perfect match. In Proceedings of the

Dissertation Award, Demonstration, and Industrial

Track at BPM 2018 co-located with 16th International

Conference on Business Process Management (BPM

2018), Sydney, Australia, September 9-14, 2018, vol-

ume 2196 of CEUR Workshop Proceedings, pages

124–131. CEUR-WS.org.

Hayes, T. R. and Henderson, J. M. (2021). Looking for

semantic similarity: what a vector-space model of

semantics can tell us about attention in real-world

scenes. Psychological Science, 32(8).

Jr., S. B., Ceravolo, P., Damiani, E., and Tavares, G. M.

(2021). Using meta-learning to recommend process

discovery methods. CoRR, abs/2103.12874.

Leemans, S. J. J., Fahland, D., and van der Aalst, W. M. P.

(2013). Discovering block-structured process models

from event logs containing infrequent behaviour. In

Business Process Management Workshops - BPM Bei-

jing, China, August 26, 2013, Revised Papers, volume

171 of Lecture Notes in Business Information Process-

ing, pages 66–78. Springer.

Leno, V., Augusto, A., Dumas, M., Rosa, M. L., Maggi,

F. M., and Polyvyanyy, A. (2020). Identifying candi-

date routines for robotic process automation from un-

segmented UI logs. In 2nd International Conference

on Process Mining, ICPM 2020, Padua, Italy, October

4-9, 2020, pages 153–160. IEEE.

Mayr, A., Herm, L., Wanner, J., and Janiesch, C. (2022).

Applications and challenges of task mining: A liter-

ature review. In 30th European Conference on Infor-

mation Systems - New Horizons in Digitally United

Societies, ECIS 2022, Timisoara, Romania, June 18-

24, 2022.

Petri, C. A. and Reisig, W. (2008). Petri net. Scholarpedia,

3(4):6477.

Qiu, X., Sun, T., Xu, Y., Shao, Y., Dai, N., and Huang,

X. (2020). Pre-trained models for natural language

processing: A survey. CoRR, abs/2003.08271.

Ribeiro, J., Carmona, J., Misir, M., and Sebag, M. (2014). A

recommender system for process discovery. In Busi-

ness Process Management - 12th International Con-

ference, BPM 2014, Haifa, Israel, September 7-11,

2014. Proceedings, volume 8659 of Lecture Notes in

Computer Science, pages 67–83. Springer.

Rozinat, A. and van der Aalst, W. M. P. (2008). Confor-

mance checking of processes based on monitoring real

behavior. Inf. Syst., 33(1):64–95.

Schonenberg, H., Weber, B., Dongen, B. v., and Aalst, W.

v. d. (2008). Supporting ﬂexible processes through

recommendations based on history. In International

Conference on Business Process Management, pages

51–66. Springer.

Seeliger, A., Nolle, T., and Mühlhäuser, M. (2018). Pro-

cess explorer: an interactive visual recommendation

system for process mining. In KDD Workshop on In-

teractive Data Exploration and Analytics.

Sroka, M. and Fani Sani, M. (2022). 50 business assign-

ments log.

Syed, R., Suriadi, S., Adams, M., Bandara, W., Leemans,

S. J., Ouyang, C., ter Hofstede, A. H., van de Weerd,

I., Wynn, M. T., and Reijers, H. A. (2020). Robotic

process automation: contemporary themes and chal-

lenges. Computers in Industry, 115:103162.

Terragni, A. and Hassani, M. (2018). Analyzing customer

journey with process mining: From discovery to rec-

ommendations. In 2018 IEEE 6th International Con-

ference on Future Internet of Things and Cloud (Fi-

Cloud), pages 224–229. IEEE.

van der Aalst, W. M. P. (2016). Process Mining - Data

Science in Action, Second Edition. Springer Berlin

Heidelberg.

Behavioral Recommender System for Process Automation Steps

261

van der Aalst, W. M. P., Rubin, V. A., Verbeek, H. M. W.,

van Dongen, B. F., Kindler, E., and Günther, C. W.

(2010). Process mining: a two-step approach to bal-

ance between underﬁtting and overﬁtting. Softw. Syst.

Model., 9(1):87–111.

Wang, J., Wong, R. K., Ding, J., Guo, Q., and Wen, L.

(2012). On recommendation of process mining algo-

rithms. In Goble, C. A., Chen, P. P., and Zhang, J.,

editors, 2012 IEEE 19th International Conference on

Web Services, Honolulu, HI, USA, June 24-29, 2012.

IEEE Computer Society.

Wanner, J., Hofmann, A., Fischer, M., Imgrund, F., Ja-

niesch, C., and Geyer-Klingeberg, J. (2019). Pro-

cess selection in RPA projects - towards a quantiﬁ-

able method of decision making. In Proceedings of

the 40th International Conference on Information Sys-

tems, ICIS, Munich, Germany, 2019. Association for

Information Systems.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

262