W.S. Yeung, A.T. Khader (Malaysia), and J. Korczak (France)
Project Management, Early Warning, Data Mining, Classifi cation
In recent years, the idea of early warning has been designed in project management to identify, analyse and forewarn user of potential problems. Many provide early warning facility, but none can determine early warning according to the ur gency of a task. This is deemed important because ignorance of critical task will lead to project failure. In addition, none provides learning ability that help project management appli cation to understand how the existing data behave and there fore prevent future risks. This paper addresses such prob lems by studying the feasibility of detecting early warning in project management using classification approach. This study assumes that historical data are available. Hence, classifica tion learning algorithms can be used. Several algorithms, such as Multi-Layer Perceptrons (MLP), Support Vector Machines (SVM), Sparse Network of Winnows (SNoW), and Decision Trees (DT) are used in experiments. After carefully tuning the hyper-parameters, MLP seems to outperform the rest of the algorithms, in terms of accuracy in evaluating warnings. In this study, we found that the historical data contains some regular patterns and that such regularity was successfully cap tured by MLP. On top of using the crisp output of MLP to determine the urgency level, we examined also the case of us ing raw output in order to answer the question: "Among tasks labeled as very critical, which one should be given more pri ority?". Although this is technically feasible, there is no way to evaluate the performance since the priority information is simply unavailable in the current data set. Nevertheless, there are strong indications that classification approach is applica ble on the early warning problem.
Important Links:
Go Back
