Repository files navigation

PyDREAM is a Python implementation of the Decay Replay Mining (DREAM) approach and the corresponding predictive algorithms (NAP and NAPr) similar to the ones described in the paperDecay Replay Mining to Predict Next Process Events and is based onPM4Py. The original Java implementation used for benchmarking can be foundhere. There exists also a ProM pluginhere.

Please seeexample.py for an end-to-end example.

PyDREAM requires an event log and a corresponding PNML Petri net file, both imported through PM4Py.

frompm4py.objects.log.importer.xesimportimporterasxes_importerfrompm4py.objects.petri.importerimportimporteraspnml_importerlog=xes_importer.apply('YOUR_EVENTLOG.xes')net,initial_marking,_=pnml_importer.apply("YOURPETRINET.pnml")

The event log must be wrapped into a PyDREAM LogWrapper instance.

frompydream.LogWrapperimportLogWrapperfrompm4py.objects.log.importer.xesimportimporterasxes_importerlog=xes_importer.apply('YOUR_EVENTLOG.xes')log_wrapper=LogWrapper(log)

If you plan on using resources, for example to train a NAPr model, provide the relevant resource identifiers.

log_wrapper=LogWrapper(log,resources=["IDENTIFIER"])

The loaded Petri net can be enhanced as described in the paper as described subsequently. AnEnhancedPN instance will automatically detect if a givenLogWrapper objects encompasses resources.

frompydream.EnhancedPNimportEnhancedPNenhanced_pn=EnhancedPN(net,initial_marking)enhanced_pn.enhance(log_wrapper)enhanced_pn.saveToFile("YOURENHANCEDPN.json")

Timed State Samples through Decay Replay can be obtained by replaying an event log wrapped into anLogWrapper instance on the enhanced Petri net. The function returns two values.First, the resulting Timed State Samples in JSON format that can be saved to file directly. Second, a list of TimedStateSample instances that can be used for further processing. The Timed State Samples will contain resource counters if a givenLogWrapper objects encompasses resources.

importjsonfrompydream.LogWrapperimportLogWrapperfrompydream.EnhancedPNimportEnhancedPNfrompm4py.objects.petri.importerimportimporteraspnml_importerfrompm4py.objects.log.importer.xesimportimporterasxes_importerlog=xes_importer.apply('YOUR_EVENTLOG.xes')log_wrapper=LogWrapper(log)net,initial_marking,_=pnml_importer.import_net("YOURPETRINET.pnml")enhanced_pn=EnhancedPN(net,initial_marking,decay_function_file="YOURENHANCEDPN.json")tss_json,tss_objs=enhanced_pn.decay_replay(log_wrapper=log_wrapper)withopen("timedstatesamples.json",'w')asf:json.dump(tss_json,f)

To replay also resources, calldecay_replay() by listing all resources identifier that should be considered.

log_wrapper=LogWrapper(log,resources=["IDENTIFIER"])tss_json,tss_objs=enhanced_pn.decay_replay(log_wrapper=log_wrapper,resources=["IDENTIFIER"])

ANAP orNAPr predictor can be trained in the following way.

frompydream.predictive.nap.NAPimportNAPalgo=NAP(tss_train_file="timedstatesamples.json",tss_test_file="timedstatesamples.json",options={"n_epochs" :100})algo.train(checkpoint_path="model-path",name="MODEL-NAME",save_results=True)

The corresponding model will be stored automatically based on the providedcheckpoint_path andname parameters. The implemented options include:

• "seed" : int
• "n_epochs" : str
• "n_batch_size" : int
• "dropout_rate" : float
• "eval_size" : float
• "activation_function" : str

One can load a trained model from file in the following way.

algo=NAP()algo.loadModel(path="model-path",name="MODEL-NAME")

The following function predicts the next events by returning the raw NAP output and the event names in string format.

frompydream.util.TimedStateSamplesimportloadTimedStateSamplestss_loaded_objs=loadTimedStateSamples("timedstatesamples.json")nap_out,string_out=algo.predict(tss_loaded_objs)

PyDREAM is developed for Python 3.7 and is based on PM4Py v2.1.2. NAP and NAPr require tensorflow 2.4.0. The full list of requirements can be found inrequirements.txt.

This code is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the corresponding license for more details.

@article{theis2019decay,  title={Decay Replay Mining to Predict Next Process Events},  author={Theis, Julian and Darabi, Houshang},  journal={IEEE Access},  volume={7},  pages={119787--119803},  year={2019},  publisher={IEEE}}