Диссертация (1137084), страница 36
Текст из файла (страница 36)
These results can not be obtained using the configuration used in previoussections.Besides, in some of these large case, a prototype implementation of the modular repairtechnique is also failed to construct the result hours. We have no repaired models. These casesindicated with R-F (repair fail).It can be easily concluded from Table 4.13, that the greedy repair of large models using ILPminer is practically impossible.
Note that re-discovery of models with a large number of activities158Table 4.13: Process model repair of larger models159is also computationally hard problem. In particular, re-discovery completely failed in the case ofthe model LMx8.What is good, naive and improved techniques are able to successfully repair all models fromthe experimental set. In all cases, a relatively small fragment of a model touched by the repair.Consider concrete examples in more detail.Figures 4.72 and 4.73 show models LMx2-BL and LMx2-BNL. The first of them has two localsmall inconsistencies. The second one contains a single non-local inconsistency.Figure 4.72: Model LMx2-BLFigure 4.73: Model LMx2-BNLThe model re-discovered from scratch using Inductive miner from the event log 3 is shown inFigure 4.74.
As in previous cases, this model significantly differs from the model LMx2-CM. Usually,the larger an initial model is, the more differences will be in the model re-discovered from an eventlog generated using the initial model.Figure 4.74: Model re-discovered from the event log 3 using Inductive minerIn the remainder, we will not show all the repaired models, but only some of them. Forexample, Figure 4.75 shows the model LMx2-BL repaired using the improved modular techniquewith Inductive miner.
As expected, the technique repaired two small enlarged fragments whichare highlighted in the figure with green colour.Figure 4.75: Model LMx2-BL repaired using the improved technique with Inductive minerYet another three models LMx4-BL, LMx4-BNL-1, LMx4-BNL-2 are shown in figures 4.76, 4.77,and 4.78. Each of these models is constructed of 4 LM2 models. The first of them contains four smalllocal inconsistencies. The second one has one non-local inconsistency. The third model containstwo inconsistencies: a local and a non-local ones. These inconsistencies are highlighted with red.160Figure 4.76: Model LMx4-BLFigure 4.77: Model LMx4-BNL-1Figure 4.78: Model LMx4-BNL-2Consider how the model LMx4-BNL-2 can be repaired using the improved technique. Figure 4.79shows the result of such a repair.
This model contains inconsistencies of two types: local and nonlocal.Three patch-ups are made in this model. As it was discussed earlier, the improved techniquerepairs local inconsistencies preserving model precision, and repairs non-local inconsistenciesreducing the precision. In the example, one may see that one of patch-ups (highlighted with3 in Figure 4.79) repairs the local inconsistency. Two remainder changed fragments (1 and 2in Figure 4.79) repair the non-local inconsistency.
Unfortunately, fragments 1 and 2 containtransitions without outgoing and incoming arcs correspondingly. Thus, precision is reduced inthis model. However, the larger a model is, the weaker an influence of each small change on thetotal value of the model.Figure 4.79: Model LMx4-BNL-2 repaired using the improved technique with Inductive minerThe largest model with eight local inconsistencies LMx8-BL is shown in Figure 4.80. Theimproved technique with both Inductive and ILP miners successfully reconstructed the initialcorrect model LMx8-CM.
The naive technique replaced 8 unfitting fragments which all contain24 nodes. The improved technique enlarged these 8 sub-nets with neighbour sub-nets. Thisenlargement increased a total number of nodes in replaced fragments to 64 nodes. Repairtechniques constructed similar models using Inductive and ILP miners.Figure 4.80: Model LMx8-BLFigure 4.81: Model LMx8-BNLThe model LMx8-BNL with a single non-local inconsistency in shown in Figure 4.81. Thisinconsistency influence on 2 of 8 LM2 models in one of parallel branches. The greedy repair issuitable for the non-local cases of such type.161Figure 4.82 shows how the model LMx8-BNL is repaired using the greedy repair techniquewith Inductive miner.
Replaced fragment contains 262 nodes of total 1062 nodes in the initialmodel. It is highlighted with a greed contour. In the re-discovered fragment, all attributes ofa model discovered by Inductive miner are present. In particular, it contains many additionalsilent transitions, choices with loops through silent transitions, and flower fragments.
However, allchanges are made in a quarter of the whole model, while the other part of it preserved.Figure 4.82: Model LMx8-BNL repaired using the greedy technique with Inductive minerThe greedy technique with ILP miner has failed to repair this model LMx8-BNL. Recall, that there-discovery from scratch also have failed in this case. Thus, there are specific high-scale cases whenthe re-discovery is infeasible while the modular repair still works.
That is so if all inconsistenciescover only a part of a model.4.4ConclusionsThis chapter considered the experimental evaluation of the modular repair technique thathas been presented in Chapter 2. Section 4.1 described the prototype implementation that weused to experiment with the technique. An experimental design with initial data are discussedin Section 4.2. Results of conducted experiments with artificial event logs and process modelsreflecting main process patterns are presented in Section 4.3.Experiments are separated into three groups.Firstly, the naive and the improved techniques are evaluated on models with localinconsistencies, for which the modular technique has been made.
Obtained results show theapplicability of both techniques, and higher effectiveness of the improved technique suitable torepair the model fitness while preserving its structure and keeping it precise.Repairs of the ProM 6 Model Repair plug-in are also presented to compare them with resultsof the modular technique. These results show that the Model Repair plug-in changes models moresignificantly than the modular technique when repairing models with local inconsistencies to theperfect fitness.
This leads to less precise models.Secondly, the modular technique applied to samples with the non-local cases. Experimentsshow, that the naive and the improved techniques repair such models in terms of fitness, but makethem less precise. More appropriate for such cases is the greedy repair technique.162Results of the ProM 6 Model Repair plug-in are better when repairing the non-local cases.Still, this approach significantly changes repaired models, but the greedy technique also changes abigger part of a model. Nevertheless, the Model Repair plug-in changes all parts of a model whenrepairing it.
At the same time, the greedy repair covers a part of a model that contain reasons ofinconsistencies, i.e. unfitting fragments. Thus, when unfitting fragments are close to each other,the greedy repair changes the model less than when such fragments are far from each other.Thirdly, results of experiments with larger models are shown. They have been conducted toevaluate scalability of the modular repair technique. The experiments illustrate the applicabilityof the naive and the improved techniques.
An efficiency of the greedy approach depends on asize of the fragment-to-repair constructed by the iterative procedure. When the whole processmodel needs repairing, the greedy technique works slowly. However, experiments show, that thereare large workflow models with non-local inconsistencies which can be repaired using the greedyapproach, but can not be re-discovered using the same discovery algorithm on the same hardware.In this chapter, we evaluated the general modular repair scheme from Chapter 2 onlywith limited configuration. In particular, the maximal decomposition joint with Inductive andILP miners has been evaluated. There are many other possible configurations which has notbeen implemented and evaluated yet. Other existing valid decompositions can be used. Forexample, SESE and Passage decompositions considered in Section 2.6 satisfy the validity criteria.Algorithms to construct these decompositions should be incorporated into the modular techniqueand evaluated.
This is one of future work directions, all of which will be listed in the conclusionsof this thesis.163ConclusionsThis chapter concludes the thesis, and consists of two sections. Main contributions are discussedin the first section.
The second one describes the open issues, and possible directions to continuethe project.Contributions of the ThesisThis thesis addressees the problem of process model repair with respect to a behaviour recordedin event logs. Namely, how to repair a workflow net so that it will perfectly fit the event log whilepreserving model’s structure, and keeping the model as precise as it has been before the repair.Process model repair techniques to solve this problem are presented in the thesis.Chapter 1 of this thesis describes the basic definitions and propositions which are used in thethesis.
In this chapter we also consider related work on model repair within the broader subjectof process mining.Process model repair techniques are presented in Chapter 2. All techniques are based on thegeneral modular scheme that can be further developed in the future. The improved technique (seeSection 2.4) is most appropriate for local inconsistencies in a model, while non-local inconsistenciescan be repaired using the greedy technique (see Section 2.5).The proposed repair methods have been evaluated using the sample data that contains typicalprocess modelling constructs. Chapter 3 describes the algorithms for generating artificial event logsfrom process models which are employed in experiments. The results of experimental evaluationare shown and discussed in Chapter 4.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
