Optimization-Based Business Process Model Matching
This is a modal window.
Das Video konnte nicht geladen werden, da entweder ein Server- oder Netzwerkfehler auftrat oder das Format nicht unterstützt wird.
Formale Metadaten
| Titel |
| |
| Serientitel | ||
| Anzahl der Teile | 30 | |
| Autor | ||
| Lizenz | CC-Namensnennung - keine kommerzielle Nutzung 4.0 International: Sie dürfen das Werk bzw. den Inhalt zu jedem legalen und nicht-kommerziellen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen. | |
| Identifikatoren | 10.5446/53691 (DOI) | |
| Herausgeber | ||
| Erscheinungsjahr | ||
| Sprache |
Inhaltliche Metadaten
| Fachgebiet | ||
| Genre | ||
| Abstract |
| |
| Schlagwörter |
11
00:27
Computeranimation
02:00
Computeranimation
04:37
Flussdiagramm
11:16
Computeranimation
15:46
Computeranimation
Transkript: Englisch(automatisch erzeugt)
00:00
Welcome to everyone. So my name is Sarah and I'm participating this conference from every day at an university from Germany. And I'm a postdoc and a group leader at the chair of the process and details as a chair, and it is led by Professor we find that as and today I'm going to just present our recent work about the optimization based business process model matching. Okay, so let's start.
00:27
I would like to first give out a small overview about the problem statement. So how it looks like and within the next 15 minutes, what kind of topic you're going to be familiar with. And I would like to just start this start with two examples. So when we have two business
00:46
process models, how the question will be then have to determine if any two models will be then matching. That means if they really match to each other, or to which extent it may be then matched. So this is actually the main question that we're going to examine today. And here the main thing is that we have here just the.
01:10
Alrighty, this is the first process model, and this is the second process model, and we would like to just have a look at the properties of these process models in order to just attain the matching right and what kind of approach will be then here used.
01:26
The approach is going to be then the first one is the optimization based matching approach. That means we're going to use an optimization problem in order to find the maximum matching.
01:40
And how are we going to do this? The idea will be then to have a look at the label and behavior information of both process models. So this is going to be the main idea that I'm trying to just give you some overview about in the upcoming slides. So just a very few words about the latest works, how the latest work looks like
02:08
with respect to the behavioral information and label information, which you can find in the literature. So I just need to point out that there are a few different areas when you talk about the matching approach, right? So, for example, you see in the literature that there are numerous label matching approach.
02:27
That means you have the activities and as I have already shown on the previous slide, then you have the activity labels. That means for each activity, you have a specific activity label, and these labels are compared with each other by using some similarity functions.
02:48
And this is one very famous and very common, let's say, field where you can use, which you can use for matching approach. And the other one would be, for example, having a look at the behavioral information.
03:03
That means when you have a process model, so what is the behavior of the activities? That means which activity is coming after which one or not necessarily after which one and what kind of, for example, concurrence information is there or is there anything, for example, like the causality relationship among the activities that you see in the process model and so on.
03:25
So this is going to be another field where you have the behavioral information, right? So here in this approach, before I just continue with the working mechanism of this approach, I would like to just mention that we have chosen these three related because they are the most interesting approach, which are directly related to our setting.
03:51
So the first one is the bi-simulation-based matching approach, which is a directly behavioral matching approach. That means if you just take two process models, then bi-simulation is played over there so that you can find a specific approach as a value then.
04:06
And the second one would be the micrologic network matching approach, where you have the behavioral matching again, and there are two different possibilities that they evaluated later on with two different labels and narrative functions. And then we have the bigger words, process model matching approach, which is also very famous.
04:25
Probably everybody knows about that in that field. And this is going to be a baseline for us because it's a legal matching approach. That means we are going to use these three then later on for our evaluation. So what about the approach itself? So the main idea is here that we have that we just need to consider the behavioral information.
04:48
So this is one aspect what we are considering here. And the other thing is that we are also using the label information. That means we are trying to make an optimization problem where we have two different parts and we will use the label information
05:07
and we will just compare them with each other from one, for example, specific activity to label in the from the second process model. And as our activity, for example, and the labels will be then compared with each other. So this is actually the one side and the other side.
05:24
We will use the behavioral information. That means we will have a look at the the orders of the activities in the first process model and the second process model. And we are just going to compare this kind of relationships with each other. So that means this is the second step and it's going to be a global step for us. And how do we do that?
05:45
We do that just by using an optimization problem that is an integer linear program, ILP. And I will not go into details due to the time limitations today, but you can find all the details in every theoretical aspect in the paper.
06:00
So what you really need to know after hearing this presentation is the first one, we maximize the label similarities among the process activities. Right. This is the first important fact. And the second one is behavioral similarity between the process models. That means when using this integer in our program, we maximize two parts,
06:25
label similarities among the process activities and the second one is the behavioral similarity. OK, so these are the two important facts that you really need to know about. And then how we do that, I mean, it's a very conceptual level.
06:40
That means that we're interested in a local local level that is labeled similarities because why are they local levels? Because we just directly examined the activities. Right. This is the activity level. That means it's a local level for us. But as a global level, that means we are interested in the behavior of the whole process model where we have all the relationships among the activities.
07:03
That means which one is coming after which one or what kind of concurrence relationships we have there. Or if it's kind of, for example, eventually fellows or always or sometimes relationships and so on. That means we just collect all this kind of behavioral profiles and we use all things in the literature, what has been already published.
07:24
And we just use these profiles in order to specify our behavioral aspects of each process model. And this is going to be a global level for us. OK, so far, so good. So let's go continue with the next slide.
07:41
And I'd like to now give you an overview about what does it mean? Just consider the label similarities. So just look at these two process models just in order to make it was really a tangible example for you to understand. So this is the first one, I mean the red one, and the second one is the green one. This is the second process model.
08:03
And we are just trying to have a look at the matching, the best matching with respect to the label similarities, right? So what does it mean? For example, in the first process model, you have the activity purchase of application and the second one is going to be the app purchase.
08:21
Right? These two activities are found to be very similar to each other and matching in terms of the label similarity values, right? And we go on like the virtual payment. So virtual payment, when we look for all the label similarities, we see the actual virtual payment activity will be then the best matching for that guy here, right?
08:45
And then when we continue the internal transaction, you see here that is going to be then directly, for example, either this transaction or this transaction activity in the second model, that would be then directly the best matching.
09:01
And credit card payment will be then matched to choose credit card payment and the external transaction, which is highlighted in red, will be then matched to the transaction in the other process model. So this is directly performed over various different, I mean, various generative functions. And we just use here, for example, linear similarity functions so that we can find some similarities among the label similarities.
09:28
OK, so what about the behavioral similarities? I have already talked about, you know, the order of the activities. So it's very important for us to know about, for example, what kind of activity will be
09:41
performed in which order when you compare it with the other ones within a specific process model. For example, when you have a virtual payment, right? After that, you have a specific transaction, that means an internal transaction, for example, here. And this is the order. That means you do not have first the transaction and then the virtual payment, but you have first the virtual payment and then the internal transaction.
10:02
Or you may also find other examples here, like the application purchase activity. This is always performed before any transactions here. So that means your virtual payment and the credit card payment, this is there. I mean, you just need to just choose one of them, right?
10:22
So this kind of relationships, they will be then stored in our system, I mean, in our optimization problems, so that we can just be aware of the behavior of the process model with respect to all the activities, how they relate to each other. And when we look at the second model here, you'll see that we have the choose virtual payment and then the transaction.
10:43
So we have also a specific ordering of the activities with respect to the one above. And when we compare exactly these four transitions, I mean, the activities with each other, we directly compare this relationship
11:01
that is directly for this relationship, for example, in this case, with each other, and it is perfectly matching, right? This is an example for you to understand how it looks like when we talk about the behavioral similarities. OK, so I hope it is clear. Then I would like to just say a few words about the experimental results, what kind of results we had.
11:25
And I will go into details and the questions at the end of the presentation. So for the experiments, we had a look at the data sets directly from the process model matching contests from 2015. And there were many three interesting data sets, the university admission processes, the birth
11:45
registration processes, and I mean, that was collected over a few hospitals and management processes. So for the university admission processes, you have different universities and mainly the same concepts where you have different models.
12:05
And when you just have a look at the characteristics, where basically then you'll see that we have some model pairs like 36 model pairs in each data set. And we have different, for example, number of transitions, the maximum number of transitions, especially when
12:20
you look at the university asset data sets, you have really a lot of transitions as maximum. And in average, you'll see that the university has actually more transitions per model. But there are more other data set characteristics, which you can find also in the paper. OK, so what about the experimental results? First of all, I can directly mention that we have
12:45
a look at the precision recall and the S4 values and with respect to the precision values. So we have already had a look at different the approach, like the bigger words, the approach and bigger words optimize going to be ours.
13:05
I mean, that means the optimization based business process model matching approach. And the third one is sorry, the third one is the Marco approach. This is with a specific similarity function. And the other one is also with another
13:23
similarity function. And the last one is the isolation approach that I have already talked about. So when we look at here on the left side, we have here the macro values and the micro values that means we have two different aggregation granularities, but it doesn't really matter.
13:41
So if you do not know about them, it's not a problem. So it's just a have a look at the first positive true positives and false negatives and true negatives. I mean, this kind of all the value is just either collecting everything. I mean, just adding everything together and then just obtaining the precision recall or the S4 values.
14:05
Or you can just do that for each model comparison and then just have an average out of that. So this is just a comparison, just just a difference. I mean, so what we can see here is for just let's consider here the first left hand side here, for example, we are the second one here.
14:24
I mean, you see that the optima is going to be the second bar and we are competing and all the data stats with the existing approaches. This is what what is important for us. But for sure, there are some differences, for example, for the let's say for the birth data stats.
14:43
It is very clear that there are some with respect to the characteristics of the data that, for example, by submission approach is much better when you go further, which is back to the recall how it looks like. Again, the second one is going to be our approach and here we are outperforming the others
15:02
in the birth data stats and as well here we are just directly competing with the Markov methods. I mean, that's also very similar to each other, what we have already gathered as a value and here as well in the university data stats. So for us, it was also important to see especially the F-score values because the F-score values you have another perspective here.
15:26
And what we have already gathered is that we have seen that the optima in our approach then can outperform mainly the all related words here. And it was especially in the university data, for example, they just said it was very tangible.
15:44
Okay, so we have seen very in a crash course, the experimental results and now I'd like to just conclude the paper. So what we have already seen within this very short presentation. So we have seen that there is a novel approach, which is back to the optimization based process model matching.
16:04
And we have seen when we have two different process models that we have the activities, I mean the labels, which can be matched to each other and the optimization problem directly tries to match the best matching. I mean, find the best matching in terms of using a maximization problem.
16:23
And we have two levels, that means the label level and the behavioral similarity level. It's important for us that we just do not look at the label similarities, but we also use the behavioral information of the process models in order to be able to find a good trade off.
16:42
That means how we can just find a good process model matching. And this can be then also derived by specific weight weighting parameter, which you can also find in the optimization problem directly in the paper. So we have maximized the label similarities and also the behavioral similarities among the processes.
17:04
And what we have already seen was that also in the paper that when we use both label and behavioral information of the process models, then we attain really very good results, but not just label, but not just behavioral information. So that was very interesting to see. And yeah, there was an optimization problem that was very clear for us.
17:25
It's a maximization problem. It's an integer linear program that we have already used. And we were able to perform related work on three real world data sets with respect to the F-score value. So as a feature work, I mean, there are a lot of, a couple of stuff that we can actually really examine.
17:44
And just two of them I can just directly mention here is that the analysis of the complex correspondence is what does it mean? For example, we can directly have a look at also blocks of the process models for if they are really matching to each other. So that would be also an interesting direction. And also, as we just wanted to just
18:05
do that, but we had not really a lot of options here because of the paper limitation. So the analysis of the execution time is also very important because if you have an optimization problem, then it's clear that there are a few maybe slow processes.
18:22
So it is important to just compare that with also those existing methods. Right. Okay. So this is from my side. Thank you very much for your attention. If you have any questions, please go ahead.