Диссертация (1137084), страница 31
Текст из файла (страница 31)
In general, these settingsare not suitable for the considered case, when the structure of repaired model must be saved.Figure 4.18: Model SM1-BL-1 repaired using the Model Repair plug-in with settings alignsub-logs and compute global cost functionFigure 4.19 shows a model repaired using the combination of loop and sub-process detection.In this experiments, the plug-in with this setting constructs models which are most similar to thecorrect model SM1-CM.
However, their fitness is usually not perfect. One may easily see, that inthe model from Figure 4.19 a transition 6 still fires before 4 .132Figure 4.19: Model SM1-BL-1 repaired using the Model Repair plug-in with loop and sub-processdetectionThe seventh complex combination of settings (almost all Yes) leads to a model shown inFigure 4.20. It is simple and similar to the initial model SM1-CM, but the inconsistency (6 firesbefore 4 ) does not repaired.Figure 4.20: Model SM1-BL-1 repaired using the Model Repair plug-in with Y-Y-N-Y-Y settingsIt can be concluded, that the plug-in Model Repair with settings which guarantee perfectfitness of a repaired model changes the model significantly.
Otherwise, settings which save model’sstructure do not guarantee perfect fitness, although the repaired model is usually fits the eventlog used to repair better than the initial model. Recall, that when new activities are added to theprocess, this repair method suits well.Table 4.3 shows characteristics of SM1-BL-1 repairs using the method of D.
Fahland andW. van der Aalst [25]. Unfortunately, in many cases plug-ins failed to evaluate fitness and precisionof a repaired model. In such cases, word fail is placed in the corresponding table cell. This failindication means the fails of a conformance checking algorithm, not a repair algorithm. Fortunately,models are relatively small for a human being to evaluate them visually.Table 4.3: Characteristics of SM1-BL-1 repairs using the Model Repair plug-inNote that in general, there is no guarantee of perfect fitness, when this method is used. However,one may choose the appropriate setting using their experience or some kind of a brute force.133Model SM1-BL-2Consider another changed model SM1-BL-2 shown in Figure 4.21.
In this example, twotransitions 4 and 5 are swapped as in the model SM1-BL-1. But, these are not neighbourtransitions, 6 transition is between them. Note that in the initial model SM1-BL-1 a transition5 plays a role of AND join. It can fire strictly after firing of 4 and 6 in the top branch, andtransitions 0 , 1 , 2 , 3 in the bottom branch. The changed model can not reproduce this correctorder of process activities. Thus, inconsistency is more significant in that case. Indeed, fitness ofthe event log 1 and the changed model SM1-BL-2 is 0,89, precision can be estimated as 0,89.Figure 4.21: Changed model SM1-BL-2Figure 4.22 shows the model SM1-BL-2 repaired using the naive repair technique with Inductiveminer. A replaced fragment is larger in that case. In particular, the sub-nets with transitions 3 , 4 ,5 , and 6 have fitness with related sub-logs lower than 1.
These sub-nets have been re-discovered,and then composed. One may note that two additional silent transitions are added to the model.These transitions reduce model precision. Fitness between 1 and the repaired model is 1. Thefragment replaced by repair contains 9 nodes of 40. Precision calculation plug-in failed to processthis model13 .
Besides, the repaired model has size of 45 nodes; 2 new silent transitions and 3 newplaces are added to the model.Figure 4.22: Model SM1-BL-2 repaired using the naive algorithm with Inductive miner13There are cases when alignment-based techniques fail to calculate results in observable time, or the alignmentcalculation procedure failed with exception. This is because the behaviour of a model is very diverse, so there is nosufficient memory for calculating fitness or precision of the model.
Many silent transitions or transitions withoutincoming arcs from restrictive places lead to such a case. Interestingly, some of models in the remainder seem tobe not very complex, but existing conformance checking tools failed to deal with them.134Figure 4.23 shows the model SM1-BL-2 repaired using the naive repair technique with ILPminer. The repaired model differs from the model repaired using Inductive miner (see Figure 4.22).Of most importance is that one of re-discovered fragments is disconnected model.
That is why 3has no outgoing arcs. The model perfectly fits 1 , but its precision is 0,50. The repaired modelcontains 39 nodes; one place is removed during repair. Total 9 nodes are touched.Figure 4.23: Model SM1-BL-2 repaired using the naive algorithm with ILP minerConsider results of the improved repair technique.
Changed fragment is larger in that case. Itcontains 15 nodes, 5 of which are borderline transitions. Inductive miner adds silent transitionsand places to the fragment. Thus, the size of repaired model (see Figure 4.24) is 47 nodes.Figure 4.24: Model SM1-BL-2 repaired using the improved algorithm with Inductive minerFigure 4.25 shows how model SM1-BL-2 is repaired using the improved repair technique withILP miner.
This model is almost the same as SM1-CM, there is only one additional place in it whichdoes not change the behaviour.Figure 4.25: Model SM1-BL-2 repaired using the improved algorithm with ILP minerBoth models repaired using the improved repair techniques perfectly fit 1 , and their precisionis 0,91. Thus, both models repaired by the improved technique are as precise as the initial modelSM1-CM.Consider how ProM 6 Model Repair plug-in is able to repair the model SM1-BL-2. Perfectlyrepaired models are complex and have unnecessary transitions and places.
Table 4.4 summarizes135Table 4.4: Characteristics of SM1-BL-2 repairs using the Model Repair plug-inthe characteristics of repaired models. Note that perfectly fitting models are constructed usingcombinations 1, 2, 4, and 5. Consider these models.Figure 4.26 shows a model repaired using the 1 combination. Figure 4.27 shows a model repairedusing the 2 combination, when loops are detected.
Infrequent nodes are removed during repairfrom the model shown in Figure 4.28 (the 4 combination). The result of using the 5 combinationis shown in Figure 4.29.Figure 4.26: Model SM1-BL-2 repaired using the Model Repair plug-in with N-N-N-N-N settingsNote that all models have many additional transitions and places. Most of added transitionsare silent. Besides, some of existing process activities are represented by a number of transitionswith the same label in repaired model.
In particular, the model from Figure 4.27 has 6 transitionslabelled with 4 (transitions which are added by the repair are shown as 4 +), and 2 5 transitions.The same transition copy is performed in all four repaired models.Figure 4.27: Model SM1-BL-2 repaired using the Model Repair plug-in with loop detectionThe most simple model is constructed using the 5 combination (see Figure 4.29). However,even this model has little in common with the initial model SM1-CM. The only part of the modelstructurally preserved that is far from the swapped transitions 4 and 5 which effect inconsistencies.Moreover, all these models are less precise than the initial model, and the models repaired usingmodular repair technique. Note that the re-discovery using ILP miner works better (see Figure 4.9)in that particular case.136Figure 4.28: Model SM1-BL-2 repaired using the Model Repair plug-in with infrequent nodesdeletionModels from Figure 4.26 and Figure 4.29 contain redundant silent transitions.
For example, inFigure 4.29 such transitions are connected to the places 1 and 5 . These transitions describeno useful behaviour. Thus, there is no harm to remove them.Figure 4.29: Model SM1-BL-2 repaired using the Model Repair plug-in with settings alignsub-logs and compute global cost functionRepairs using combinations 3, 6, and 7 are simpler and closer to the initial model, but thefitness is less than 1, as for the model SM1-BL-1.Model LM2-BL-1Consider experiments with a larger model LM2-CM.
Three models with local changes areprepared: LM2-BL-1, LM2-BL-2, and LM2-BL-3. Figure 4.30 shows the first of them, LM2-BL-1.Fitness between this model and the event log 2 is 0,98, precision can be estimated as 0,59.Note that the initial model LM2-CM has the following characteristics: (2 , LM2-CM) = 1,(2 , LM2-CM) = 0,59.Figure 4.30: Changed model LM2-BL-1Results of naive repair are shown in Figure 4.31. Note that two sub-nets have inconsistenciesin that case. Replaced sub-nets are highlighted using a green colour in figures. Repaired modelsobtained using both Inductive or ILP miners have the same structure, thus we show only one ofthem.
Fitness between the repaired model and 2 is 1, whereas precision is 0,28. Both changedfragments contain 7 nodes (5 transitions and 2 places) of total 133 nodes in the whole model.137Note that this model is significantly more similar with an initial model LM2-CM, than models rediscovered using Inductive and ILP miners (see Figure 4.7 and Figure 4.8). This is the reason toapply the model repair instead of the re-discovery.Figure 4.31: Model LM2-BL-1 repaired using the naive algorithm with Inductive or ILP minersIn that case, the improved repair technique with Inductive and ILP miners return the sameresult, as naive repair.
Figure 4.32 shows the model. All unfitting sub-nets and their neighbourfitting sub-nets are joined into a single fragment. This fragment consists of 12 nodes of 133 andhas two borderline transitions. Actually, the improved technique reconstructs the initial modelLM2-CM, and have the same characteristics.Figure 4.32: Model LM2-BL-1 repaired using the improved algorithm with Inductive or ILPminersModels based on the LM2-CM are relatively large. Because of that less settings combinationsof the Model Repair plug-in produced perfectly fitting models.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
