Диссертация (1137084), страница 32
Текст из файла (страница 32)
Consider repaired modelcharacteristics in Table 4.5. Only three combinations repaired models perfectly.Table 4.5: Characteristics of LM2-BL-1 repairs using the Model Repair plug-inFigure 4.33 shows how model LM2-BL-1 is repaired using the combination 2. Surprisingly,loop detection setting works relatively fine in this case, and allows for repairing model whereasits structure is mostly preserved. However, the modular repair technique (see Figure 4.32) evenreconstructs the initial model.Other perfectly fitting repairs using Model Repair plug-in have much more differences with theinitial model.
Indeed, compare models depicted in Figure 4.34 or in Figure 4.35 with the initialmodel LM2-CM shown in Figure 4.6.138Figure 4.33: Model LM2-BL-1 repaired using the Model Repair plug-in with loop detectionA large number of changes is obvious. Repaired models are full of additional silent transitions.Moreover, a lot of existing transitions copied during the repair. Note how almost the whole modelis changed.Figure 4.34: Model LM2-BL-1 repaired using the Model Repair plug-in with infrequent nodesdeletionThis complex structure of a repaired model with lots of silent transitions leads to a verydiverse model behaviour. In turn, such a behaviour troubles conformance checking algorithms.Model simplification techniques described in Section 1.3.7 can help in this case.Figure 4.35: Model LM2-BL-1 repaired using the Model Repair plug-in with settings alignsub-logs and compute global cost functionOther repairs of the Model Repair plug-in do not guarantee perfect fitness of the repairedmodel, and , actually, do not repair inconsistencies of LM2-BL-1.
Thus, they are not presentedhere, but can be found at the page of this project.Further two more local model repair examples are considered.Model LM2-BL-2Figure 4.36 shows the example, when inconsistency can be repaired by the replacement of thesingle fragment. This model is almost perfectly fits the event log 2 . In particular, fitness is 0,99,and precision can be estimated as 0,59.In this case, the naive repair technique constructs the same model using Inductive and ILPminers. Figure 4.37 shows this model.
Its fitness with 2 is 1, and precision is 0,38. A single139Figure 4.36: Changed model LM2-BL-2fragment is replaced which contains 3 nodes: 2 borderline transitions and 1 place. Note howsignificantly such a small change influence model’s precision (from 0,59 to 0,38).Figure 4.37: Model LM2-BL-2 repaired using the naive algorithm with Inductive or ILP minersInterestingly, models repaired using the improved technique with Inductive and ILP are thesame.
Fugure 4.38 shows the repaired model. Two neighbours are joined to the single unfittingsub-net. Thus, the replaced fragment contains more nodes: 8 of 133. This fragment is connected tothe whole model via three border transitions: 11 , 14 , and 15 . Again, the improved techniquereconstructs the initial model LM2-CM, and have the same characteristics.Figure 4.38: Model LM2-BL-2 repaired using the improved algorithm with Inductive or ILPminersThis is the most simple inconsistency of the considered earlier. It is especially insignificantwith respect to a model’s size. Thus, the Model Repair plug-in is able to perfectly repair such aproblem with many different settings.
Nevertheless, settings combinations 6 and 7 do not provideuser with a repaired model.In this case, we show only two examples of repair to keep a reader awake. Figure 4.39 showsthe result of repair using the most simple setting (all no). Note that this model is relativelysimple in comparison with other repairs by the Model Repair plug-in, but it is more complexin comparison with the model repaired using the modular repair technique (see Figure 4.38).Actually, the improved modular repair technique is able to reconstruct the initial model LM2-CMin that case (see Fugure 4.38).Figure 4.40 shows the repaired model constructed using the settings Align Sub-logs andCompute Global Cost Function.
This model perfectly fits 2 but is very complex, as in previouscases. Note that all models repaired using the Model Repair plug-in are less precise than the initial140Figure 4.39: Model LM2-BL-2 repaired using the Model Repair plug-in with N-N-N-N-N settingsFigure 4.40: Model LM2-BL-2 repaired using the Model Repair plug-in with settings alignsub-logs and compute global cost functionmodel LM2-CM and a model before repair (LM2-BL-2 in this case). The same is correct for modelsdiscovered from scratch based on 2 (see Figure 4.9 and Figure 4.10). A model repaired using thenaive algorithm repairs model is significantly less precise as well.
At the same time, the improvedrepair algorithm reconstructs the initial model LM2-CM.Results of the Model Repair plug-in for this model are shown in Table 4.6.Table 4.6: Characteristics of LM2-BL-2 repairs using the Model Repair plug-inModel LM2-BL-3Local inconsistencies may be spread in different parts of the model. Still, naive and improvedtechniques are able to repair them since they replace all unfitting fragments of a model bycorresponding re-discovered fitting fragments.
The locality of an inconsistency means that allcausing reasons are close to each other in the model, whereas multiple inconsistencies with differentreasons can be in the model. Non-local inconsistencies are due to the reasons which can not becovered by a single fragment of the decomposition (no matter what valid decomposition is used).Thus, their repair using naive or improved techniques changes fragment borders and reducesprecision of the result model.Consider yet another example of local model repair. Figure 4.41 shows a model LM2-BL-3 withtwo pairs of swapped transitions.
Two transitions 12 and 16 are positioned close to the sourceplace of workflow net, whereas transitions 05 and 06 are close to the sink place. Both pairsare highlighted in Figure 4.41 with red colour. Note that each inconsistency is local, as swapped141transitions in each pair can be covered with an enlarged fragment of the improved repair technique.Fitness between the event log 2 and the initial model LM2-BL-3 is 0,97, and precision is 0,59.Figure 4.41: Changed model LM2-BL-3In that case, repairs using Inductive and ILP miners construct the same models. The resultof naive repair using both mining algorithms is shown in Figure 4.42.
Note that there are threeunfitting sub-nets which are replaced. The first pair (transitions 05 and 06 ) are in one sub-net.The second pair induces two unfitting sub-nets. All replaced sub-nets are highlighted in Figure 4.42with green colour. Total 10 nodes of 133 are located in these three sub-nets. This model perfectlyfits the event log 2 .
A model’s precision is very low: 0,23.Figure 4.42: Model LM2-BL-3 repaired using the naive algorithm with Inductive or ILP minersHere again, the improved repair technique produces the same model using Inductive and ILPminers. Figure 4.43 shows this model. This model has the same characteristics as the initial modelLM2-CM. Its fitness with 2 is 1, and precision is 0,59. All fragments touched by the repair contain21 nodes of total 133 nodes in the initial model LM2-BL-3.Figure 4.43: Model LM2-BL-3 repaired using the improved algorithm with Inductive or ILPminersThe result model of the repair using this technique is almost the initial model LM2-CM.
Theonly difference is that the single place (which is 12 ) is added to the model. Thus, the repairedmodel contains 134 nodes. Note that this place is totally redundant, and has no influence onthe behaviour of the model. Figure 4.44 shows an fragment of the model from Figure 4.43. Theredundant place 12 is highlighted with the orange coloured box. This place is added by a modeldiscovery algorithm because the transition 02 always fires before the 04 .
Events 07 , 03 , and08 are filtered out from the sub-log related to the repaired fragment, and the corresponding modelpart is not involved in repair. Thus there should be a place that will represent a causal dependencybetween 02 and 04 . This place is 12 . When the fragments are composed, 12 becomes142redundant as in the initial model already there are parts which represent causal dependencybetween 02 and 04 .
Such places can be removed from the result model using process modelsimplification techniques similar to the one described in Section 1.3.7.Figure 4.44: One of two replaced fragments with a redundant transitionThe Model Repair plug-in significantly changes the model when repairs many localinconsistencies in one model. Interestingly, this fact is not reflected in the results of numericalevaluation.
They are shown in Table 4.7. This may be, when structurally more complex modelsrepresent more or less the same behaviour. Indeed, consider two following perfectly fitting modelswhich produced by the repair method.Table 4.7: Characteristics of LM2-BL-3 repairs using the Model Repair plug-inFigure 4.45 shows the model repaired using the simplest settings combination. It is easy tonote that this model structurally has little in common with the correct model LM2-CM and thebroken model LM2-BL-3.
The repair changed model completely. As in previous cases, the modelcontains many silent transitions and copies of transitions with the same label. This fact troublesconformance checking, and makes the model less human-readable.Figure 4.45: Model LM2-BL-3 repaired using the Model Repair plug-in with N-N-N-N-N settings143Even more complex model is constructed using the setting combination 5. This model is shownin Figure 4.46.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
