داده کاوی جریان کار با InWoLvE

State of the art information systems are based on explicit process models called workflow models. Experience from industrial practice shows that the definition of workflow models is a very time consuming and error prone task. Recently, there has been an increasing interest in applying techniques from data mining and machine learning to support this task. This approach has also been termed as process or workflow mining. In this paper, we give an overview of the algorithms that were implemented within the InWoLvE workflow mining system, we summarize the most important results of their experimental evaluation and we present the experiences that were made in the first industrial application of InWoLvE.

State of the art information systems are based on explicit process models called workflow models. These models are interpreted by one or more workflow engines to drive the execution of business processes within or across several enterprises. Experience from industrial practice shows that the definition of workflow models is a very time consuming and error prone task. In depth knowledge of the business process and the ability to represent this knowledge using a formal workflow modelling language are needed for this task. Recently, there has been an increasing interest in applying techniques from data mining and machine learning to support this task [1], [2], [3], [4], [5], [6], [7], [8] and [9]. This approach has also been termed as process or workflow mining. The basic idea of the workflow mining approach is to collect traces of workflow executions and to derive a workflow model from these observations. This is useful for example if some information system supporting the process, that logs all relevant events, is already in place before the workflow model is defined. Furthermore workflow mining techniques and advanced workflow technology, which is moving towards more operational flexibility [10], [11], [12] and [13], enable an evolutionary approach to the development of workflow applications, where an initially roughly defined and informal or semi-formal workflow model is iteratively refined and formalized. In this paper, we give an overview of the algorithms that were implemented within the InWoLvE workflow mining system, we summarize the most important results of their experimental evaluation and we present the experiences that were made in the first industrial application of InWoLvE. The remainder of this paper is organized as follows. Section 2 defines the most important terms used throughout this paper. Section 3 formalizes the workflow mining problem, it defines problem classes and it gives an overview of the induction and transformation algorithms used within the workflow mining system InWoLvE. In Section 4, we describe the InWoLvE prototype and we summarize the most important results of its experimental evaluation. The experiences we have made in the first industrial application of InWoLvE are covered by Section 5. In Section 6, we discuss related work and finally in Section 7 we summarize the main conclusions and give an outlook on our future work.

7. Conclusions and future work We have described the most important algorithms that were implemented in the workflow mining system InWoLvE. InWoLvE solves the workflow mining task in two steps. In the first step it creates a stochastic activity graph from the example set and in the second step it transforms this stochastic activity graph into a well-defined workflow model. The experiments we have described show that InWoLvE is applicable for a wide range of workflow models. We have also presented the experiences made during the first successful application of InWoLvE within a workflow project at DaimlerChrysler. Our experiments also showed us a number of directions for the improvement of InWoLvE, which we will be working on in the future: • The strategy for removing repeated nodes can be improved. We have been thinking about an operator for changing the enumeration of repeated activities. An opposite search strategy may also help solving this problem. • The technique for the detection of dependencies can also be improved. Possible extensions include the use of observed dataflows for dependency detection. • The experiments have shown that workflow mining is an iterative and interactive process. This needs to be supported by InWoLvE much better, by showing intermediate results and allowing the workflow miner to take decisions. • Further experiences with real workflow traces need to be collected.