یک رویکرد نیمه خودکار برای انتساب کارمندان جریان کار
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|21811||2008||14 صفحه PDF||سفارش دهید||8553 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Computers in Industry, Volume 59, Issue 5, May 2008, Pages 463–476
Staff assignment is of great importance for workflow management systems. In many workflow applications, staff assignment is still performed manually. In this paper, we present a semi-automatic approach intended to reduce the number of manual staff assignment. Our approach applies a machine learning algorithm to the workflow event log to learn various kinds of activities that each actor undertakes. When staff assignment is needed, the classifiers generated by the machine learning technique suggest a suitable actor to undertake the specified activities. With experiments on three enterprises, our approach achieved a fairly accurate recommendation.
In the context of workflow, staff assignment serves for the purpose of specifying the relationship between activities and resources . It ensures that the operation of a workflow conforms to its intended design principles and operates as efficiently and deterministically as possible. In most cases, staff assignment is performed at workflow build-time stage to restrict the range of resources that can undertake an activity, usually by means of the “role” concept. At run-time stage, workflow engine automatically assigns the work to all the resources with that role or to specific resource using some simple mechanisms such as queue lengths or round-robin, etc. . However, in many real situations, such simple run-time work allocation mechanisms are not sufficient for organizations to correctly assign work to resources. Consider, for example, an engineering design process in a car manufacturing enterprise we investigated, a typical part design activity is defined to be undertaken by the resources who belong to the designer role, but the actual designer who is responsible for the design can only be specified when a concrete requirement arrives. Because this task is directly related to the manufacturing and eventually determines the quality of the final product, it is unlikely to let designers arbitrarily accept the work items. Therefore, run-time staff assignment is needed. Usually it is performed manually by workflow initiators or monitors. To our knowledge, such manual staff assignment occurs frequently in manufacturing enterprises, especially for those important tasks in business processes. Although there are various mechanisms of staff assignment in the literature of workflow , ,  and , few of them focus on actively recommending resources at run-time, especially using the workflow history information. In Ref. , Muehlen envisioned that the workflow history information could be used to improve workflow run-time resource allocation. Russell et al. introduced a pattern (R-HBA), which offers or allocates work items to resources on the basis of their previous execution history, but none of the investigated workflow systems provide direct support for this pattern . Ly and Rinderle et al. proposed a method to derive staff assignment rules from event logs that mainly aims at facilitating the assignment at build-time stage  and . In this paper, we present an approach intended to reduce the amount of manual staff assignment performed at workflow run-time instantiation and execution stages. Our approach applies a machine learning algorithm to the workflow event log in order to learn various kinds of activities that each actor undertakes. When staff assignment is needed, the classifiers generated by the machine learning technique suggest a suitable actor to undertake the specified activities. Our approach requires an enterprise's workflow system to have had an event log for some period of time and the corresponding workflow models, so that the patterns of who executes what kinds of activities can be learned. Using our approach, we have been able to correctly suggest appropriate actors to undertake the activities with overall prediction accuracies of 82.88%, 79.48% and 79.44%, respectively in three vehicle manufacturing enterprises. This paper makes two contributions: firstly, it presents an approach for helping automate workflow staff assignment in workflow management systems; secondly, it evaluates the applicability of a machine learning approach for staff assignment using real world datasets. This paper is organized as follows: we begin with presenting some background information about the workflow event log and general information about three enterprises (Section 2). Given this background, we describe our semi-automated approach for staff assignment (Section 3) and present the results of applying our approach on real world datasets (Section 4). We then discuss some possible improvements (Section 5). Related works about workflow resources allocation are discussed in Section 6. Finally, we summarize the paper (Section 7).
نتیجه گیری انگلیسی
Despite the results we report in this paper, it is still reasonable for us to believe that there is plenty of space for improvement. In this section, we discuss some possible directions. Machine learning algorithms generally produce better results with more data available from which to learn. Therefore, prediction accuracy can be further improved by incorporating other data. In our opinion, there are two promising kinds of data. One kind of data is actor's expertise information and their social relationship  and . The approach presented in Ref.  is an example of using such kind of information. The other kind of data is workflow application data. John and Langley work  can be viewed as an example. For the first kind of data, its applicability depends on how well the expertise of actors and their relationship information can be expressed by the system, which inevitably leads to the discussion of the sophisticated capability model for people and organizations. Although there are many kinds of models that are presented in the literature, most of them are actually difficult to implement for current workflow systems. Besides, the cost of practicing such kinds of models is rather expensive, hence the availability of this part of information is a big issue. For the second kind of data, the workflow application data are very promising to be incorporated into learning, at least, its availability is much better than the expertise information. The obstacle faced is that data are usually application context dependent, which means that different enterprise's data contain different information. Therefore, special approaches are needed to extract useful feature information, for example, using process data warehouse . Nevertheless, the feedback from enterprise users has also revealed the fact that people often make their decisions based on the documents that are processed by the workflow. This fact motivates us to carry out further investigation. In the real situation, the workflow event log is not available all at a time, however, the approach we present in this paper trains the classifier using a batched set of data. Alternatively an incremental algorithm could be used whereby instances are provided one at a time to the classifier and the classifier updates itself accordingly . This incremental approach will be more suitable than the batched approach especially for those enterprises whose workflow system is intensively used. Moreover, the approach we have presented in this paper only recommends potential actors one at a time. As a matter of fact, groups of actors often work on similar kinds of activities. Therefore, it might be more helpful to recommend a small list of potential actors rather than just one. However, determining the best group of actors is not as easy as it seems to be. Further experiments and development of the learning approach are needed.